Mon, 12 Mar 2012 15:28:07 -0700
7152957: VM crashes with assert(false) failed: bad AD file
Reviewed-by: kvn, never
Contributed-by: nils.eliasson@oracle.com
1 /*
2 * Copyright (c) 2008, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "interpreter/interpreter.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "prims/methodHandles.hpp"
30 #define __ _masm->
32 #ifdef PRODUCT
33 #define BLOCK_COMMENT(str) /* nothing */
34 #else
35 #define BLOCK_COMMENT(str) __ block_comment(str)
36 #endif
38 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
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 }
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);
57 return (address) me;
58 }
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");
65 // Fill in the real end_address:
66 __ align(wordSize);
67 me->set_end_address(__ pc());
69 return me;
70 }
72 // stack walking support
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 }
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);
89 // pick up the argument type descriptor:
90 Thread* thread = Thread::current();
91 Handle cookie(thread, f->compute_saved_args_layout(true, true));
93 // process fixed part
94 blk->do_oop((oop*)f->saved_target_addr());
95 blk->do_oop((oop*)f->saved_args_layout_addr());
97 // process variable arguments:
98 if (cookie.is_null()) return; // no arguments to describe
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 }
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
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 }
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;
162 address start = __ pc();
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
172 // Return values are in registers.
173 // L1_continuation contains a cleanup continuation we must return
174 // to.
176 (*bounce_offset) = __ pc() - start;
177 BLOCK_COMMENT("ricochet_blob.bounce");
179 if (VerifyMethodHandles) RicochetFrame::verify_clean(_masm);
180 trace_method_handle(_masm, "return/ricochet_blob.bounce");
182 __ JMP(L1_continuation, 0);
183 __ delayed()->nop();
184 __ illtrap(0);
186 DEBUG_ONLY(__ set(MAGIC_NUMBER_2, G0));
188 (*exception_offset) = __ pc() - start;
189 BLOCK_COMMENT("ricochet_blob.exception");
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);
198 // Take down the frame.
200 // Cf. InterpreterMacroAssembler::remove_activation.
201 leave_ricochet_frame(_masm, /*recv_reg=*/ noreg, I5_savedSP, I7);
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());
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);
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 }
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());
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 }
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 }
270 // Emit code to verify that FP is pointing at a valid ricochet frame.
271 #ifndef PRODUCT
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 #endif
280 #ifdef ASSERT
281 void MethodHandles::RicochetFrame::verify_clean(MacroAssembler* _masm) {
282 // The stack should look like this:
283 // ... keep1 | dest=42 | keep2 | magic | handler | magic | recursive args | [RF]
284 // Check various invariants.
286 Register O7_temp = O7, O5_temp = O5;
288 Label L_ok_1, L_ok_2, L_ok_3, L_ok_4;
289 BLOCK_COMMENT("verify_clean {");
290 // Magic numbers must check out:
291 __ set((int32_t) MAGIC_NUMBER_1, O7_temp);
292 __ cmp_and_br_short(O7_temp, L0_magic_number_1, Assembler::equal, Assembler::pt, L_ok_1);
293 __ stop("damaged ricochet frame: MAGIC_NUMBER_1 not found");
295 __ BIND(L_ok_1);
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_and_brx_short(L4_saved_args_base, FP_temp, Assembler::greaterEqualUnsigned, Assembler::pt, L_ok_2);
305 __ stop("damaged ricochet frame: L4 < FP");
307 __ BIND(L_ok_2);
308 // Disable until we decide on it's fate
309 // __ sub(L4_saved_args_base, UNREASONABLE_STACK_MOVE * Interpreter::stackElementSize, O7_temp);
310 // __ cmp(O7_temp, FP_temp);
311 // __ br(Assembler::lessEqualUnsigned, false, Assembler::pt, L_ok_3);
312 // __ delayed()->nop();
313 // __ stop("damaged ricochet frame: (L4 - UNREASONABLE_STACK_MOVE) > FP");
315 __ BIND(L_ok_3);
316 extract_conversion_dest_type(_masm, L5_conversion, O7_temp);
317 __ cmp_and_br_short(O7_temp, T_VOID, Assembler::equal, Assembler::pt, L_ok_4);
318 extract_conversion_vminfo(_masm, L5_conversion, O5_temp);
319 __ ld_ptr(L4_saved_args_base, __ argument_offset(O5_temp, O5_temp), O7_temp);
320 assert(Assembler::is_simm13(RETURN_VALUE_PLACEHOLDER), "must be simm13");
321 __ cmp_and_brx_short(O7_temp, (int32_t) RETURN_VALUE_PLACEHOLDER, Assembler::equal, Assembler::pt, L_ok_4);
322 __ stop("damaged ricochet frame: RETURN_VALUE_PLACEHOLDER not found");
323 __ BIND(L_ok_4);
324 BLOCK_COMMENT("} verify_clean");
325 }
326 #endif //ASSERT
328 void MethodHandles::load_klass_from_Class(MacroAssembler* _masm, Register klass_reg, Register temp_reg, Register temp2_reg) {
329 if (VerifyMethodHandles)
330 verify_klass(_masm, klass_reg, SystemDictionaryHandles::Class_klass(), temp_reg, temp2_reg,
331 "AMH argument is a Class");
332 __ load_heap_oop(Address(klass_reg, java_lang_Class::klass_offset_in_bytes()), klass_reg);
333 }
335 void MethodHandles::load_conversion_vminfo(MacroAssembler* _masm, Address conversion_field_addr, Register reg) {
336 assert(CONV_VMINFO_SHIFT == 0, "preshifted");
337 assert(CONV_VMINFO_MASK == right_n_bits(BitsPerByte), "else change type of following load");
338 __ ldub(conversion_field_addr.plus_disp(BytesPerInt - 1), reg);
339 }
341 void MethodHandles::extract_conversion_vminfo(MacroAssembler* _masm, Register conversion_field_reg, Register reg) {
342 assert(CONV_VMINFO_SHIFT == 0, "preshifted");
343 __ and3(conversion_field_reg, CONV_VMINFO_MASK, reg);
344 }
346 void MethodHandles::extract_conversion_dest_type(MacroAssembler* _masm, Register conversion_field_reg, Register reg) {
347 __ srl(conversion_field_reg, CONV_DEST_TYPE_SHIFT, reg);
348 __ and3(reg, 0x0F, reg);
349 }
351 void MethodHandles::load_stack_move(MacroAssembler* _masm,
352 Address G3_amh_conversion,
353 Register stack_move_reg) {
354 BLOCK_COMMENT("load_stack_move {");
355 __ ldsw(G3_amh_conversion, stack_move_reg);
356 __ sra(stack_move_reg, CONV_STACK_MOVE_SHIFT, stack_move_reg);
357 #ifdef ASSERT
358 if (VerifyMethodHandles) {
359 Label L_ok, L_bad;
360 int32_t stack_move_limit = 0x0800; // extra-large
361 __ cmp_and_br_short(stack_move_reg, stack_move_limit, Assembler::greaterEqual, Assembler::pn, L_bad);
362 __ cmp(stack_move_reg, -stack_move_limit);
363 __ br(Assembler::greater, false, Assembler::pt, L_ok);
364 __ delayed()->nop();
365 __ BIND(L_bad);
366 __ stop("load_stack_move of garbage value");
367 __ BIND(L_ok);
368 }
369 #endif
370 BLOCK_COMMENT("} load_stack_move");
371 }
373 #ifdef ASSERT
374 void MethodHandles::RicochetFrame::verify() const {
375 assert(magic_number_1() == MAGIC_NUMBER_1, "");
376 if (!Universe::heap()->is_gc_active()) {
377 if (saved_args_layout() != NULL) {
378 assert(saved_args_layout()->is_method(), "must be valid oop");
379 }
380 if (saved_target() != NULL) {
381 assert(java_lang_invoke_MethodHandle::is_instance(saved_target()), "checking frame value");
382 }
383 }
384 int conv_op = adapter_conversion_op(conversion());
385 assert(conv_op == java_lang_invoke_AdapterMethodHandle::OP_COLLECT_ARGS ||
386 conv_op == java_lang_invoke_AdapterMethodHandle::OP_FOLD_ARGS ||
387 conv_op == java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_REF,
388 "must be a sane conversion");
389 if (has_return_value_slot()) {
390 assert(*return_value_slot_addr() == RETURN_VALUE_PLACEHOLDER, "");
391 }
392 }
394 void MethodHandles::verify_argslot(MacroAssembler* _masm, Register argslot_reg, Register temp_reg, const char* error_message) {
395 // Verify that argslot lies within (Gargs, FP].
396 Label L_ok, L_bad;
397 BLOCK_COMMENT("verify_argslot {");
398 __ cmp_and_brx_short(Gargs, argslot_reg, Assembler::greaterUnsigned, Assembler::pn, L_bad);
399 __ add(FP, STACK_BIAS, temp_reg); // STACK_BIAS is zero on !_LP64
400 __ cmp_and_brx_short(argslot_reg, temp_reg, Assembler::lessEqualUnsigned, Assembler::pt, L_ok);
401 __ BIND(L_bad);
402 __ stop(error_message);
403 __ BIND(L_ok);
404 BLOCK_COMMENT("} verify_argslot");
405 }
407 void MethodHandles::verify_argslots(MacroAssembler* _masm,
408 RegisterOrConstant arg_slots,
409 Register arg_slot_base_reg,
410 Register temp_reg,
411 Register temp2_reg,
412 bool negate_argslots,
413 const char* error_message) {
414 // Verify that [argslot..argslot+size) lies within (Gargs, FP).
415 Label L_ok, L_bad;
416 BLOCK_COMMENT("verify_argslots {");
417 if (negate_argslots) {
418 if (arg_slots.is_constant()) {
419 arg_slots = -1 * arg_slots.as_constant();
420 } else {
421 __ neg(arg_slots.as_register(), temp_reg);
422 arg_slots = temp_reg;
423 }
424 }
425 __ add(arg_slot_base_reg, __ argument_offset(arg_slots, temp_reg), temp_reg);
426 __ add(FP, STACK_BIAS, temp2_reg); // STACK_BIAS is zero on !_LP64
427 __ cmp_and_brx_short(temp_reg, temp2_reg, Assembler::greaterUnsigned, Assembler::pn, L_bad);
428 // Gargs points to the first word so adjust by BytesPerWord
429 __ add(arg_slot_base_reg, BytesPerWord, temp_reg);
430 __ cmp_and_brx_short(Gargs, temp_reg, Assembler::lessEqualUnsigned, Assembler::pt, L_ok);
431 __ BIND(L_bad);
432 __ stop(error_message);
433 __ BIND(L_ok);
434 BLOCK_COMMENT("} verify_argslots");
435 }
437 // Make sure that arg_slots has the same sign as the given direction.
438 // If (and only if) arg_slots is a assembly-time constant, also allow it to be zero.
439 void MethodHandles::verify_stack_move(MacroAssembler* _masm,
440 RegisterOrConstant arg_slots, int direction) {
441 enum { UNREASONABLE_STACK_MOVE = 256 * 4 }; // limit of 255 arguments
442 bool allow_zero = arg_slots.is_constant();
443 if (direction == 0) { direction = +1; allow_zero = true; }
444 assert(stack_move_unit() == -1, "else add extra checks here");
445 if (arg_slots.is_register()) {
446 Label L_ok, L_bad;
447 BLOCK_COMMENT("verify_stack_move {");
448 // __ btst(-stack_move_unit() - 1, arg_slots.as_register()); // no need
449 // __ br(Assembler::notZero, false, Assembler::pn, L_bad);
450 // __ delayed()->nop();
451 __ cmp(arg_slots.as_register(), (int32_t) NULL_WORD);
452 if (direction > 0) {
453 __ br(allow_zero ? Assembler::less : Assembler::lessEqual, false, Assembler::pn, L_bad);
454 __ delayed()->nop();
455 __ cmp(arg_slots.as_register(), (int32_t) UNREASONABLE_STACK_MOVE);
456 __ br(Assembler::less, false, Assembler::pn, L_ok);
457 __ delayed()->nop();
458 } else {
459 __ br(allow_zero ? Assembler::greater : Assembler::greaterEqual, false, Assembler::pn, L_bad);
460 __ delayed()->nop();
461 __ cmp(arg_slots.as_register(), (int32_t) -UNREASONABLE_STACK_MOVE);
462 __ br(Assembler::greater, false, Assembler::pn, L_ok);
463 __ delayed()->nop();
464 }
465 __ BIND(L_bad);
466 if (direction > 0)
467 __ stop("assert arg_slots > 0");
468 else
469 __ stop("assert arg_slots < 0");
470 __ BIND(L_ok);
471 BLOCK_COMMENT("} verify_stack_move");
472 } else {
473 intptr_t size = arg_slots.as_constant();
474 if (direction < 0) size = -size;
475 assert(size >= 0, "correct direction of constant move");
476 assert(size < UNREASONABLE_STACK_MOVE, "reasonable size of constant move");
477 }
478 }
480 void MethodHandles::verify_klass(MacroAssembler* _masm,
481 Register obj_reg, KlassHandle klass,
482 Register temp_reg, Register temp2_reg,
483 const char* error_message) {
484 oop* klass_addr = klass.raw_value();
485 assert(klass_addr >= SystemDictionaryHandles::Object_klass().raw_value() &&
486 klass_addr <= SystemDictionaryHandles::Long_klass().raw_value(),
487 "must be one of the SystemDictionaryHandles");
488 Label L_ok, L_bad;
489 BLOCK_COMMENT("verify_klass {");
490 __ verify_oop(obj_reg);
491 __ br_null_short(obj_reg, Assembler::pn, L_bad);
492 __ load_klass(obj_reg, temp_reg);
493 __ set(ExternalAddress(klass_addr), temp2_reg);
494 __ ld_ptr(Address(temp2_reg, 0), temp2_reg);
495 __ cmp_and_brx_short(temp_reg, temp2_reg, Assembler::equal, Assembler::pt, L_ok);
496 intptr_t super_check_offset = klass->super_check_offset();
497 __ ld_ptr(Address(temp_reg, super_check_offset), temp_reg);
498 __ set(ExternalAddress(klass_addr), temp2_reg);
499 __ ld_ptr(Address(temp2_reg, 0), temp2_reg);
500 __ cmp_and_brx_short(temp_reg, temp2_reg, Assembler::equal, Assembler::pt, L_ok);
501 __ BIND(L_bad);
502 __ stop(error_message);
503 __ BIND(L_ok);
504 BLOCK_COMMENT("} verify_klass");
505 }
506 #endif // ASSERT
509 void MethodHandles::jump_from_method_handle(MacroAssembler* _masm, Register method, Register target, Register temp) {
510 assert(method == G5_method, "interpreter calling convention");
511 __ verify_oop(method);
512 __ ld_ptr(G5_method, in_bytes(methodOopDesc::from_interpreted_offset()), target);
513 if (JvmtiExport::can_post_interpreter_events()) {
514 // JVMTI events, such as single-stepping, are implemented partly by avoiding running
515 // compiled code in threads for which the event is enabled. Check here for
516 // interp_only_mode if these events CAN be enabled.
517 __ verify_thread();
518 Label skip_compiled_code;
520 const Address interp_only(G2_thread, JavaThread::interp_only_mode_offset());
521 __ ld(interp_only, temp);
522 __ tst(temp);
523 __ br(Assembler::notZero, true, Assembler::pn, skip_compiled_code);
524 __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), target);
525 __ bind(skip_compiled_code);
526 }
527 __ jmp(target, 0);
528 __ delayed()->nop();
529 }
532 // Code generation
533 address MethodHandles::generate_method_handle_interpreter_entry(MacroAssembler* _masm) {
534 // I5_savedSP/O5_savedSP: sender SP (must preserve)
535 // G4 (Gargs): incoming argument list (must preserve)
536 // G5_method: invoke methodOop
537 // G3_method_handle: receiver method handle (must load from sp[MethodTypeForm.vmslots])
538 // O0, O1, O2, O3, O4: garbage temps, blown away
539 Register O0_mtype = O0;
540 Register O1_scratch = O1;
541 Register O2_scratch = O2;
542 Register O3_scratch = O3;
543 Register O4_argslot = O4;
544 Register O4_argbase = O4;
546 // emit WrongMethodType path first, to enable back-branch from main path
547 Label wrong_method_type;
548 __ bind(wrong_method_type);
549 Label invoke_generic_slow_path;
550 assert(methodOopDesc::intrinsic_id_size_in_bytes() == sizeof(u1), "");;
551 __ ldub(Address(G5_method, methodOopDesc::intrinsic_id_offset_in_bytes()), O1_scratch);
552 __ cmp(O1_scratch, (int) vmIntrinsics::_invokeExact);
553 __ brx(Assembler::notEqual, false, Assembler::pt, invoke_generic_slow_path);
554 __ delayed()->nop();
555 __ mov(O0_mtype, G5_method_type); // required by throw_WrongMethodType
556 __ mov(G3_method_handle, G3_method_handle); // already in this register
557 // O0 will be filled in with JavaThread in stub
558 __ jump_to(AddressLiteral(StubRoutines::throw_WrongMethodTypeException_entry()), O3_scratch);
559 __ delayed()->nop();
561 // here's where control starts out:
562 __ align(CodeEntryAlignment);
563 address entry_point = __ pc();
565 // fetch the MethodType from the method handle
566 // FIXME: Interpreter should transmit pre-popped stack pointer, to locate base of arg list.
567 // This would simplify several touchy bits of code.
568 // See 6984712: JSR 292 method handle calls need a clean argument base pointer
569 {
570 Register tem = G5_method;
571 for (jint* pchase = methodOopDesc::method_type_offsets_chain(); (*pchase) != -1; pchase++) {
572 __ ld_ptr(Address(tem, *pchase), O0_mtype);
573 tem = O0_mtype; // in case there is another indirection
574 }
575 }
577 // given the MethodType, find out where the MH argument is buried
578 __ load_heap_oop(Address(O0_mtype, __ delayed_value(java_lang_invoke_MethodType::form_offset_in_bytes, O1_scratch)), O4_argslot);
579 __ ldsw( Address(O4_argslot, __ delayed_value(java_lang_invoke_MethodTypeForm::vmslots_offset_in_bytes, O1_scratch)), O4_argslot);
580 __ add(__ argument_address(O4_argslot, O4_argslot, 1), O4_argbase);
581 // Note: argument_address uses its input as a scratch register!
582 Address mh_receiver_slot_addr(O4_argbase, -Interpreter::stackElementSize);
583 __ ld_ptr(mh_receiver_slot_addr, G3_method_handle);
585 trace_method_handle(_masm, "invokeExact");
587 __ check_method_handle_type(O0_mtype, G3_method_handle, O1_scratch, wrong_method_type);
589 // Nobody uses the MH receiver slot after this. Make sure.
590 DEBUG_ONLY(__ set((int32_t) 0x999999, O1_scratch); __ st_ptr(O1_scratch, mh_receiver_slot_addr));
592 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
594 // for invokeGeneric (only), apply argument and result conversions on the fly
595 __ bind(invoke_generic_slow_path);
596 #ifdef ASSERT
597 if (VerifyMethodHandles) {
598 Label L;
599 __ ldub(Address(G5_method, methodOopDesc::intrinsic_id_offset_in_bytes()), O1_scratch);
600 __ cmp(O1_scratch, (int) vmIntrinsics::_invokeGeneric);
601 __ brx(Assembler::equal, false, Assembler::pt, L);
602 __ delayed()->nop();
603 __ stop("bad methodOop::intrinsic_id");
604 __ bind(L);
605 }
606 #endif //ASSERT
608 // make room on the stack for another pointer:
609 insert_arg_slots(_masm, 2 * stack_move_unit(), O4_argbase, O1_scratch, O2_scratch, O3_scratch);
610 // load up an adapter from the calling type (Java weaves this)
611 Register O2_form = O2_scratch;
612 Register O3_adapter = O3_scratch;
613 __ load_heap_oop(Address(O0_mtype, __ delayed_value(java_lang_invoke_MethodType::form_offset_in_bytes, O1_scratch)), O2_form);
614 __ load_heap_oop(Address(O2_form, __ delayed_value(java_lang_invoke_MethodTypeForm::genericInvoker_offset_in_bytes, O1_scratch)), O3_adapter);
615 __ verify_oop(O3_adapter);
616 __ st_ptr(O3_adapter, Address(O4_argbase, 1 * Interpreter::stackElementSize));
617 // As a trusted first argument, pass the type being called, so the adapter knows
618 // the actual types of the arguments and return values.
619 // (Generic invokers are shared among form-families of method-type.)
620 __ st_ptr(O0_mtype, Address(O4_argbase, 0 * Interpreter::stackElementSize));
621 // FIXME: assert that O3_adapter is of the right method-type.
622 __ mov(O3_adapter, G3_method_handle);
623 trace_method_handle(_masm, "invokeGeneric");
624 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
626 return entry_point;
627 }
629 // Workaround for C++ overloading nastiness on '0' for RegisterOrConstant.
630 static RegisterOrConstant constant(int value) {
631 return RegisterOrConstant(value);
632 }
634 static void load_vmargslot(MacroAssembler* _masm, Address vmargslot_addr, Register result) {
635 __ ldsw(vmargslot_addr, result);
636 }
638 static RegisterOrConstant adjust_SP_and_Gargs_down_by_slots(MacroAssembler* _masm,
639 RegisterOrConstant arg_slots,
640 Register temp_reg, Register temp2_reg) {
641 // Keep the stack pointer 2*wordSize aligned.
642 const int TwoWordAlignmentMask = right_n_bits(LogBytesPerWord + 1);
643 if (arg_slots.is_constant()) {
644 const int offset = arg_slots.as_constant() << LogBytesPerWord;
645 const int masked_offset = round_to(offset, 2 * BytesPerWord);
646 const int masked_offset2 = (offset + 1*BytesPerWord) & ~TwoWordAlignmentMask;
647 assert(masked_offset == masked_offset2, "must agree");
648 __ sub(Gargs, offset, Gargs);
649 __ sub(SP, masked_offset, SP );
650 return offset;
651 } else {
652 #ifdef ASSERT
653 {
654 Label L_ok;
655 __ cmp_and_br_short(arg_slots.as_register(), 0, Assembler::greaterEqual, Assembler::pt, L_ok);
656 __ stop("negative arg_slots");
657 __ bind(L_ok);
658 }
659 #endif
660 __ sll_ptr(arg_slots.as_register(), LogBytesPerWord, temp_reg);
661 __ add( temp_reg, 1*BytesPerWord, temp2_reg);
662 __ andn(temp2_reg, TwoWordAlignmentMask, temp2_reg);
663 __ sub(Gargs, temp_reg, Gargs);
664 __ sub(SP, temp2_reg, SP );
665 return temp_reg;
666 }
667 }
669 static RegisterOrConstant adjust_SP_and_Gargs_up_by_slots(MacroAssembler* _masm,
670 RegisterOrConstant arg_slots,
671 Register temp_reg, Register temp2_reg) {
672 // Keep the stack pointer 2*wordSize aligned.
673 const int TwoWordAlignmentMask = right_n_bits(LogBytesPerWord + 1);
674 if (arg_slots.is_constant()) {
675 const int offset = arg_slots.as_constant() << LogBytesPerWord;
676 const int masked_offset = offset & ~TwoWordAlignmentMask;
677 __ add(Gargs, offset, Gargs);
678 __ add(SP, masked_offset, SP );
679 return offset;
680 } else {
681 __ sll_ptr(arg_slots.as_register(), LogBytesPerWord, temp_reg);
682 __ andn(temp_reg, TwoWordAlignmentMask, temp2_reg);
683 __ add(Gargs, temp_reg, Gargs);
684 __ add(SP, temp2_reg, SP );
685 return temp_reg;
686 }
687 }
689 // Helper to insert argument slots into the stack.
690 // arg_slots must be a multiple of stack_move_unit() and < 0
691 // argslot_reg is decremented to point to the new (shifted) location of the argslot
692 // But, temp_reg ends up holding the original value of argslot_reg.
693 void MethodHandles::insert_arg_slots(MacroAssembler* _masm,
694 RegisterOrConstant arg_slots,
695 Register argslot_reg,
696 Register temp_reg, Register temp2_reg, Register temp3_reg) {
697 // allow constant zero
698 if (arg_slots.is_constant() && arg_slots.as_constant() == 0)
699 return;
701 assert_different_registers(argslot_reg, temp_reg, temp2_reg, temp3_reg,
702 (!arg_slots.is_register() ? Gargs : arg_slots.as_register()));
704 BLOCK_COMMENT("insert_arg_slots {");
705 if (VerifyMethodHandles)
706 verify_argslot(_masm, argslot_reg, temp_reg, "insertion point must fall within current frame");
707 if (VerifyMethodHandles)
708 verify_stack_move(_masm, arg_slots, -1);
710 // Make space on the stack for the inserted argument(s).
711 // Then pull down everything shallower than argslot_reg.
712 // The stacked return address gets pulled down with everything else.
713 // That is, copy [sp, argslot) downward by -size words. In pseudo-code:
714 // sp -= size;
715 // for (temp = sp + size; temp < argslot; temp++)
716 // temp[-size] = temp[0]
717 // argslot -= size;
719 // offset is temp3_reg in case of arg_slots being a register.
720 RegisterOrConstant offset = adjust_SP_and_Gargs_up_by_slots(_masm, arg_slots, temp3_reg, temp_reg);
721 __ sub(Gargs, offset, temp_reg); // source pointer for copy
723 {
724 Label loop;
725 __ BIND(loop);
726 // pull one word down each time through the loop
727 __ ld_ptr( Address(temp_reg, 0 ), temp2_reg);
728 __ st_ptr(temp2_reg, Address(temp_reg, offset) );
729 __ add(temp_reg, wordSize, temp_reg);
730 __ cmp_and_brx_short(temp_reg, argslot_reg, Assembler::lessUnsigned, Assembler::pt, loop);
731 }
733 // Now move the argslot down, to point to the opened-up space.
734 __ add(argslot_reg, offset, argslot_reg);
735 BLOCK_COMMENT("} insert_arg_slots");
736 }
739 // Helper to remove argument slots from the stack.
740 // arg_slots must be a multiple of stack_move_unit() and > 0
741 void MethodHandles::remove_arg_slots(MacroAssembler* _masm,
742 RegisterOrConstant arg_slots,
743 Register argslot_reg,
744 Register temp_reg, Register temp2_reg, Register temp3_reg) {
745 // allow constant zero
746 if (arg_slots.is_constant() && arg_slots.as_constant() == 0)
747 return;
748 assert_different_registers(argslot_reg, temp_reg, temp2_reg, temp3_reg,
749 (!arg_slots.is_register() ? Gargs : arg_slots.as_register()));
751 BLOCK_COMMENT("remove_arg_slots {");
752 if (VerifyMethodHandles)
753 verify_argslots(_masm, arg_slots, argslot_reg, temp_reg, temp2_reg, false,
754 "deleted argument(s) must fall within current frame");
755 if (VerifyMethodHandles)
756 verify_stack_move(_masm, arg_slots, +1);
758 // Pull up everything shallower than argslot.
759 // Then remove the excess space on the stack.
760 // The stacked return address gets pulled up with everything else.
761 // That is, copy [sp, argslot) upward by size words. In pseudo-code:
762 // for (temp = argslot-1; temp >= sp; --temp)
763 // temp[size] = temp[0]
764 // argslot += size;
765 // sp += size;
767 RegisterOrConstant offset = __ regcon_sll_ptr(arg_slots, LogBytesPerWord, temp3_reg);
768 __ sub(argslot_reg, wordSize, temp_reg); // source pointer for copy
770 {
771 Label L_loop;
772 __ BIND(L_loop);
773 // pull one word up each time through the loop
774 __ ld_ptr( Address(temp_reg, 0 ), temp2_reg);
775 __ st_ptr(temp2_reg, Address(temp_reg, offset) );
776 __ sub(temp_reg, wordSize, temp_reg);
777 __ cmp_and_brx_short(temp_reg, Gargs, Assembler::greaterEqualUnsigned, Assembler::pt, L_loop);
778 }
780 // And adjust the argslot address to point at the deletion point.
781 __ add(argslot_reg, offset, argslot_reg);
783 // We don't need the offset at this point anymore, just adjust SP and Gargs.
784 (void) adjust_SP_and_Gargs_up_by_slots(_masm, arg_slots, temp3_reg, temp_reg);
786 BLOCK_COMMENT("} remove_arg_slots");
787 }
789 // Helper to copy argument slots to the top of the stack.
790 // The sequence starts with argslot_reg and is counted by slot_count
791 // slot_count must be a multiple of stack_move_unit() and >= 0
792 // This function blows the temps but does not change argslot_reg.
793 void MethodHandles::push_arg_slots(MacroAssembler* _masm,
794 Register argslot_reg,
795 RegisterOrConstant slot_count,
796 Register temp_reg, Register temp2_reg) {
797 // allow constant zero
798 if (slot_count.is_constant() && slot_count.as_constant() == 0)
799 return;
800 assert_different_registers(argslot_reg, temp_reg, temp2_reg,
801 (!slot_count.is_register() ? Gargs : slot_count.as_register()),
802 SP);
803 assert(Interpreter::stackElementSize == wordSize, "else change this code");
805 BLOCK_COMMENT("push_arg_slots {");
806 if (VerifyMethodHandles)
807 verify_stack_move(_masm, slot_count, 0);
809 RegisterOrConstant offset = adjust_SP_and_Gargs_down_by_slots(_masm, slot_count, temp2_reg, temp_reg);
811 if (slot_count.is_constant()) {
812 for (int i = slot_count.as_constant() - 1; i >= 0; i--) {
813 __ ld_ptr( Address(argslot_reg, i * wordSize), temp_reg);
814 __ st_ptr(temp_reg, Address(Gargs, i * wordSize));
815 }
816 } else {
817 Label L_plural, L_loop, L_break;
818 // Emit code to dynamically check for the common cases, zero and one slot.
819 __ cmp(slot_count.as_register(), (int32_t) 1);
820 __ br(Assembler::greater, false, Assembler::pn, L_plural);
821 __ delayed()->nop();
822 __ br(Assembler::less, false, Assembler::pn, L_break);
823 __ delayed()->nop();
824 __ ld_ptr( Address(argslot_reg, 0), temp_reg);
825 __ st_ptr(temp_reg, Address(Gargs, 0));
826 __ ba_short(L_break);
827 __ BIND(L_plural);
829 // Loop for 2 or more:
830 // top = &argslot[slot_count]
831 // while (top > argslot) *(--Gargs) = *(--top)
832 Register top_reg = temp_reg;
833 __ add(argslot_reg, offset, top_reg);
834 __ add(Gargs, offset, Gargs ); // move back up again so we can go down
835 __ BIND(L_loop);
836 __ sub(top_reg, wordSize, top_reg);
837 __ sub(Gargs, wordSize, Gargs );
838 __ ld_ptr( Address(top_reg, 0), temp2_reg);
839 __ st_ptr(temp2_reg, Address(Gargs, 0));
840 __ cmp_and_brx_short(top_reg, argslot_reg, Assembler::greaterUnsigned, Assembler::pt, L_loop);
841 __ BIND(L_break);
842 }
843 BLOCK_COMMENT("} push_arg_slots");
844 }
846 // in-place movement; no change to Gargs
847 // blows temp_reg, temp2_reg
848 void MethodHandles::move_arg_slots_up(MacroAssembler* _masm,
849 Register bottom_reg, // invariant
850 Address top_addr, // can use temp_reg
851 RegisterOrConstant positive_distance_in_slots, // destroyed if register
852 Register temp_reg, Register temp2_reg) {
853 assert_different_registers(bottom_reg,
854 temp_reg, temp2_reg,
855 positive_distance_in_slots.register_or_noreg());
856 BLOCK_COMMENT("move_arg_slots_up {");
857 Label L_loop, L_break;
858 Register top_reg = temp_reg;
859 if (!top_addr.is_same_address(Address(top_reg, 0))) {
860 __ add(top_addr, top_reg);
861 }
862 // Detect empty (or broken) loop:
863 #ifdef ASSERT
864 if (VerifyMethodHandles) {
865 // Verify that &bottom < &top (non-empty interval)
866 Label L_ok, L_bad;
867 if (positive_distance_in_slots.is_register()) {
868 __ cmp(positive_distance_in_slots.as_register(), (int32_t) 0);
869 __ br(Assembler::lessEqual, false, Assembler::pn, L_bad);
870 __ delayed()->nop();
871 }
872 __ cmp_and_brx_short(bottom_reg, top_reg, Assembler::lessUnsigned, Assembler::pt, L_ok);
873 __ BIND(L_bad);
874 __ stop("valid bounds (copy up)");
875 __ BIND(L_ok);
876 }
877 #endif
878 __ cmp_and_brx_short(bottom_reg, top_reg, Assembler::greaterEqualUnsigned, Assembler::pn, L_break);
879 // work top down to bottom, copying contiguous data upwards
880 // In pseudo-code:
881 // while (--top >= bottom) *(top + distance) = *(top + 0);
882 RegisterOrConstant offset = __ argument_offset(positive_distance_in_slots, positive_distance_in_slots.register_or_noreg());
883 __ BIND(L_loop);
884 __ sub(top_reg, wordSize, top_reg);
885 __ ld_ptr( Address(top_reg, 0 ), temp2_reg);
886 __ st_ptr(temp2_reg, Address(top_reg, offset) );
887 __ cmp_and_brx_short(top_reg, bottom_reg, Assembler::greaterUnsigned, Assembler::pt, L_loop);
888 assert(Interpreter::stackElementSize == wordSize, "else change loop");
889 __ BIND(L_break);
890 BLOCK_COMMENT("} move_arg_slots_up");
891 }
893 // in-place movement; no change to rsp
894 // blows temp_reg, temp2_reg
895 void MethodHandles::move_arg_slots_down(MacroAssembler* _masm,
896 Address bottom_addr, // can use temp_reg
897 Register top_reg, // invariant
898 RegisterOrConstant negative_distance_in_slots, // destroyed if register
899 Register temp_reg, Register temp2_reg) {
900 assert_different_registers(top_reg,
901 negative_distance_in_slots.register_or_noreg(),
902 temp_reg, temp2_reg);
903 BLOCK_COMMENT("move_arg_slots_down {");
904 Label L_loop, L_break;
905 Register bottom_reg = temp_reg;
906 if (!bottom_addr.is_same_address(Address(bottom_reg, 0))) {
907 __ add(bottom_addr, bottom_reg);
908 }
909 // Detect empty (or broken) loop:
910 #ifdef ASSERT
911 assert(!negative_distance_in_slots.is_constant() || negative_distance_in_slots.as_constant() < 0, "");
912 if (VerifyMethodHandles) {
913 // Verify that &bottom < &top (non-empty interval)
914 Label L_ok, L_bad;
915 if (negative_distance_in_slots.is_register()) {
916 __ cmp(negative_distance_in_slots.as_register(), (int32_t) 0);
917 __ br(Assembler::greaterEqual, false, Assembler::pn, L_bad);
918 __ delayed()->nop();
919 }
920 __ cmp_and_brx_short(bottom_reg, top_reg, Assembler::lessUnsigned, Assembler::pt, L_ok);
921 __ BIND(L_bad);
922 __ stop("valid bounds (copy down)");
923 __ BIND(L_ok);
924 }
925 #endif
926 __ cmp_and_brx_short(bottom_reg, top_reg, Assembler::greaterEqualUnsigned, Assembler::pn, L_break);
927 // work bottom up to top, copying contiguous data downwards
928 // In pseudo-code:
929 // while (bottom < top) *(bottom - distance) = *(bottom + 0), bottom++;
930 RegisterOrConstant offset = __ argument_offset(negative_distance_in_slots, negative_distance_in_slots.register_or_noreg());
931 __ BIND(L_loop);
932 __ ld_ptr( Address(bottom_reg, 0 ), temp2_reg);
933 __ st_ptr(temp2_reg, Address(bottom_reg, offset) );
934 __ add(bottom_reg, wordSize, bottom_reg);
935 __ cmp_and_brx_short(bottom_reg, top_reg, Assembler::lessUnsigned, Assembler::pt, L_loop);
936 assert(Interpreter::stackElementSize == wordSize, "else change loop");
937 __ BIND(L_break);
938 BLOCK_COMMENT("} move_arg_slots_down");
939 }
941 // Copy from a field or array element to a stacked argument slot.
942 // is_element (ignored) says whether caller is loading an array element instead of an instance field.
943 void MethodHandles::move_typed_arg(MacroAssembler* _masm,
944 BasicType type, bool is_element,
945 Address value_src, Address slot_dest,
946 Register temp_reg) {
947 assert(!slot_dest.uses(temp_reg), "must be different register");
948 BLOCK_COMMENT(!is_element ? "move_typed_arg {" : "move_typed_arg { (array element)");
949 if (type == T_OBJECT || type == T_ARRAY) {
950 __ load_heap_oop(value_src, temp_reg);
951 __ verify_oop(temp_reg);
952 __ st_ptr(temp_reg, slot_dest);
953 } else if (type != T_VOID) {
954 int arg_size = type2aelembytes(type);
955 bool arg_is_signed = is_signed_subword_type(type);
956 int slot_size = is_subword_type(type) ? type2aelembytes(T_INT) : arg_size; // store int sub-words as int
957 __ load_sized_value( value_src, temp_reg, arg_size, arg_is_signed);
958 __ store_sized_value(temp_reg, slot_dest, slot_size );
959 }
960 BLOCK_COMMENT("} move_typed_arg");
961 }
963 // Cf. TemplateInterpreterGenerator::generate_return_entry_for and
964 // InterpreterMacroAssembler::save_return_value
965 void MethodHandles::move_return_value(MacroAssembler* _masm, BasicType type,
966 Address return_slot) {
967 BLOCK_COMMENT("move_return_value {");
968 // Look at the type and pull the value out of the corresponding register.
969 if (type == T_VOID) {
970 // nothing to do
971 } else if (type == T_OBJECT) {
972 __ verify_oop(O0);
973 __ st_ptr(O0, return_slot);
974 } else if (type == T_INT || is_subword_type(type)) {
975 int type_size = type2aelembytes(T_INT);
976 __ store_sized_value(O0, return_slot, type_size);
977 } else if (type == T_LONG) {
978 // store the value by parts
979 // Note: We assume longs are continguous (if misaligned) on the interpreter stack.
980 #if !defined(_LP64) && defined(COMPILER2)
981 __ stx(G1, return_slot);
982 #else
983 #ifdef _LP64
984 __ stx(O0, return_slot);
985 #else
986 if (return_slot.has_disp()) {
987 // The displacement is a constant
988 __ st(O0, return_slot);
989 __ st(O1, return_slot.plus_disp(Interpreter::stackElementSize));
990 } else {
991 __ std(O0, return_slot);
992 }
993 #endif
994 #endif
995 } else if (type == T_FLOAT) {
996 __ stf(FloatRegisterImpl::S, Ftos_f, return_slot);
997 } else if (type == T_DOUBLE) {
998 __ stf(FloatRegisterImpl::D, Ftos_f, return_slot);
999 } else {
1000 ShouldNotReachHere();
1001 }
1002 BLOCK_COMMENT("} move_return_value");
1003 }
1005 #ifndef PRODUCT
1006 void MethodHandles::RicochetFrame::describe(const frame* fr, FrameValues& values, int frame_no) {
1007 RicochetFrame* rf = new RicochetFrame(*fr);
1009 // ricochet slots (kept in registers for sparc)
1010 values.describe(frame_no, rf->register_addr(I5_savedSP), err_msg("exact_sender_sp reg for #%d", frame_no));
1011 values.describe(frame_no, rf->register_addr(L5_conversion), err_msg("conversion reg for #%d", frame_no));
1012 values.describe(frame_no, rf->register_addr(L4_saved_args_base), err_msg("saved_args_base reg for #%d", frame_no));
1013 values.describe(frame_no, rf->register_addr(L3_saved_args_layout), err_msg("saved_args_layout reg for #%d", frame_no));
1014 values.describe(frame_no, rf->register_addr(L2_saved_target), err_msg("saved_target reg for #%d", frame_no));
1015 values.describe(frame_no, rf->register_addr(L1_continuation), err_msg("continuation reg for #%d", frame_no));
1017 // relevant ricochet targets (in caller frame)
1018 values.describe(-1, rf->saved_args_base(), err_msg("*saved_args_base for #%d", frame_no));
1019 values.describe(-1, (intptr_t *)(STACK_BIAS+(uintptr_t)rf->exact_sender_sp()), err_msg("*exact_sender_sp+STACK_BIAS for #%d", frame_no));
1020 }
1021 #endif // ASSERT
1023 #ifndef PRODUCT
1024 extern "C" void print_method_handle(oop mh);
1025 void trace_method_handle_stub(const char* adaptername,
1026 oopDesc* mh,
1027 intptr_t* saved_sp,
1028 intptr_t* args,
1029 intptr_t* tracing_fp) {
1030 bool has_mh = (strstr(adaptername, "return/") == NULL); // return adapters don't have mh
1032 tty->print_cr("MH %s mh="INTPTR_FORMAT " saved_sp=" INTPTR_FORMAT " args=" INTPTR_FORMAT, adaptername, (intptr_t) mh, saved_sp, args);
1034 if (Verbose) {
1035 // dumping last frame with frame::describe
1037 JavaThread* p = JavaThread::active();
1039 ResourceMark rm;
1040 PRESERVE_EXCEPTION_MARK; // may not be needed by safer and unexpensive here
1041 FrameValues values;
1043 // Note: We want to allow trace_method_handle from any call site.
1044 // While trace_method_handle creates a frame, it may be entered
1045 // without a valid return PC in O7 (e.g. not just after a call).
1046 // Walking that frame could lead to failures due to that invalid PC.
1047 // => carefully detect that frame when doing the stack walking
1049 // walk up to the right frame using the "tracing_fp" argument
1050 intptr_t* cur_sp = StubRoutines::Sparc::flush_callers_register_windows_func()();
1051 frame cur_frame(cur_sp, frame::unpatchable, NULL);
1053 while (cur_frame.fp() != (intptr_t *)(STACK_BIAS+(uintptr_t)tracing_fp)) {
1054 cur_frame = os::get_sender_for_C_frame(&cur_frame);
1055 }
1057 // safely create a frame and call frame::describe
1058 intptr_t *dump_sp = cur_frame.sender_sp();
1059 intptr_t *dump_fp = cur_frame.link();
1061 bool walkable = has_mh; // whether the traced frame shoud be walkable
1063 // the sender for cur_frame is the caller of trace_method_handle
1064 if (walkable) {
1065 // The previous definition of walkable may have to be refined
1066 // if new call sites cause the next frame constructor to start
1067 // failing. Alternatively, frame constructors could be
1068 // modified to support the current or future non walkable
1069 // frames (but this is more intrusive and is not considered as
1070 // part of this RFE, which will instead use a simpler output).
1071 frame dump_frame = frame(dump_sp,
1072 cur_frame.sp(), // younger_sp
1073 false); // no adaptation
1074 dump_frame.describe(values, 1);
1075 } else {
1076 // Robust dump for frames which cannot be constructed from sp/younger_sp
1077 // Add descriptions without building a Java frame to avoid issues
1078 values.describe(-1, dump_fp, "fp for #1 <not parsed, cannot trust pc>");
1079 values.describe(-1, dump_sp, "sp");
1080 }
1082 bool has_args = has_mh; // whether Gargs is meaningful
1084 // mark args, if seems valid (may not be valid for some adapters)
1085 if (has_args) {
1086 if ((args >= dump_sp) && (args < dump_fp)) {
1087 values.describe(-1, args, "*G4_args");
1088 }
1089 }
1091 // mark saved_sp, if seems valid (may not be valid for some adapters)
1092 intptr_t *unbiased_sp = (intptr_t *)(STACK_BIAS+(uintptr_t)saved_sp);
1093 if ((unbiased_sp >= dump_sp - UNREASONABLE_STACK_MOVE) && (unbiased_sp < dump_fp)) {
1094 values.describe(-1, unbiased_sp, "*saved_sp+STACK_BIAS");
1095 }
1097 // Note: the unextended_sp may not be correct
1098 tty->print_cr(" stack layout:");
1099 values.print(p);
1100 }
1102 if (has_mh) {
1103 print_method_handle(mh);
1104 }
1105 }
1107 void MethodHandles::trace_method_handle(MacroAssembler* _masm, const char* adaptername) {
1108 if (!TraceMethodHandles) return;
1109 BLOCK_COMMENT("trace_method_handle {");
1110 // save: Gargs, O5_savedSP
1111 __ save_frame(16); // need space for saving required FPU state
1113 __ set((intptr_t) adaptername, O0);
1114 __ mov(G3_method_handle, O1);
1115 __ mov(I5_savedSP, O2);
1116 __ mov(Gargs, O3);
1117 __ mov(I6, O4); // frame identifier for safe stack walking
1119 // Save scratched registers that might be needed. Robustness is more
1120 // important than optimizing the saves for this debug only code.
1122 // save FP result, valid at some call sites (adapter_opt_return_float, ...)
1123 Address d_save(FP, -sizeof(jdouble) + STACK_BIAS);
1124 __ stf(FloatRegisterImpl::D, Ftos_d, d_save);
1125 // Safely save all globals but G2 (handled by call_VM_leaf) and G7
1126 // (OS reserved).
1127 __ mov(G3_method_handle, L3);
1128 __ mov(Gargs, L4);
1129 __ mov(G5_method_type, L5);
1130 __ mov(G6, L6);
1131 __ mov(G1, L1);
1133 __ call_VM_leaf(L2 /* for G2 */, CAST_FROM_FN_PTR(address, trace_method_handle_stub));
1135 __ mov(L3, G3_method_handle);
1136 __ mov(L4, Gargs);
1137 __ mov(L5, G5_method_type);
1138 __ mov(L6, G6);
1139 __ mov(L1, G1);
1140 __ ldf(FloatRegisterImpl::D, d_save, Ftos_d);
1142 __ restore();
1143 BLOCK_COMMENT("} trace_method_handle");
1144 }
1145 #endif // PRODUCT
1147 // which conversion op types are implemented here?
1148 int MethodHandles::adapter_conversion_ops_supported_mask() {
1149 return ((1<<java_lang_invoke_AdapterMethodHandle::OP_RETYPE_ONLY)
1150 |(1<<java_lang_invoke_AdapterMethodHandle::OP_RETYPE_RAW)
1151 |(1<<java_lang_invoke_AdapterMethodHandle::OP_CHECK_CAST)
1152 |(1<<java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_PRIM)
1153 |(1<<java_lang_invoke_AdapterMethodHandle::OP_REF_TO_PRIM)
1154 // OP_PRIM_TO_REF is below...
1155 |(1<<java_lang_invoke_AdapterMethodHandle::OP_SWAP_ARGS)
1156 |(1<<java_lang_invoke_AdapterMethodHandle::OP_ROT_ARGS)
1157 |(1<<java_lang_invoke_AdapterMethodHandle::OP_DUP_ARGS)
1158 |(1<<java_lang_invoke_AdapterMethodHandle::OP_DROP_ARGS)
1159 // OP_COLLECT_ARGS is below...
1160 |(1<<java_lang_invoke_AdapterMethodHandle::OP_SPREAD_ARGS)
1161 |(
1162 java_lang_invoke_MethodTypeForm::vmlayout_offset_in_bytes() <= 0 ? 0 :
1163 ((1<<java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_REF)
1164 |(1<<java_lang_invoke_AdapterMethodHandle::OP_COLLECT_ARGS)
1165 |(1<<java_lang_invoke_AdapterMethodHandle::OP_FOLD_ARGS)
1166 )
1167 )
1168 );
1169 }
1171 //------------------------------------------------------------------------------
1172 // MethodHandles::generate_method_handle_stub
1173 //
1174 // Generate an "entry" field for a method handle.
1175 // This determines how the method handle will respond to calls.
1176 void MethodHandles::generate_method_handle_stub(MacroAssembler* _masm, MethodHandles::EntryKind ek) {
1177 MethodHandles::EntryKind ek_orig = ek_original_kind(ek);
1179 // Here is the register state during an interpreted call,
1180 // as set up by generate_method_handle_interpreter_entry():
1181 // - G5: garbage temp (was MethodHandle.invoke methodOop, unused)
1182 // - G3: receiver method handle
1183 // - O5_savedSP: sender SP (must preserve)
1185 const Register O0_scratch = O0;
1186 const Register O1_scratch = O1;
1187 const Register O2_scratch = O2;
1188 const Register O3_scratch = O3;
1189 const Register O4_scratch = O4;
1190 const Register G5_scratch = G5;
1192 // Often used names:
1193 const Register O0_argslot = O0;
1195 // Argument registers for _raise_exception:
1196 const Register O0_code = O0;
1197 const Register O1_actual = O1;
1198 const Register O2_required = O2;
1200 guarantee(java_lang_invoke_MethodHandle::vmentry_offset_in_bytes() != 0, "must have offsets");
1202 // Some handy addresses:
1203 Address G3_mh_vmtarget( G3_method_handle, java_lang_invoke_MethodHandle::vmtarget_offset_in_bytes());
1205 Address G3_dmh_vmindex( G3_method_handle, java_lang_invoke_DirectMethodHandle::vmindex_offset_in_bytes());
1207 Address G3_bmh_vmargslot( G3_method_handle, java_lang_invoke_BoundMethodHandle::vmargslot_offset_in_bytes());
1208 Address G3_bmh_argument( G3_method_handle, java_lang_invoke_BoundMethodHandle::argument_offset_in_bytes());
1210 Address G3_amh_vmargslot( G3_method_handle, java_lang_invoke_AdapterMethodHandle::vmargslot_offset_in_bytes());
1211 Address G3_amh_argument ( G3_method_handle, java_lang_invoke_AdapterMethodHandle::argument_offset_in_bytes());
1212 Address G3_amh_conversion(G3_method_handle, java_lang_invoke_AdapterMethodHandle::conversion_offset_in_bytes());
1214 const int java_mirror_offset = in_bytes(Klass::java_mirror_offset());
1216 if (have_entry(ek)) {
1217 __ nop(); // empty stubs make SG sick
1218 return;
1219 }
1221 address interp_entry = __ pc();
1223 trace_method_handle(_masm, entry_name(ek));
1225 BLOCK_COMMENT(err_msg("Entry %s {", entry_name(ek)));
1227 switch ((int) ek) {
1228 case _raise_exception:
1229 {
1230 // Not a real MH entry, but rather shared code for raising an
1231 // exception. For sharing purposes the arguments are passed into registers
1232 // and then placed in the intepreter calling convention here.
1233 assert(raise_exception_method(), "must be set");
1234 assert(raise_exception_method()->from_compiled_entry(), "method must be linked");
1236 __ set(AddressLiteral((address) &_raise_exception_method), G5_method);
1237 __ ld_ptr(Address(G5_method, 0), G5_method);
1239 const int jobject_oop_offset = 0;
1240 __ ld_ptr(Address(G5_method, jobject_oop_offset), G5_method);
1242 adjust_SP_and_Gargs_down_by_slots(_masm, 3, noreg, noreg);
1244 __ st (O0_code, __ argument_address(constant(2), noreg, 0));
1245 __ st_ptr(O1_actual, __ argument_address(constant(1), noreg, 0));
1246 __ st_ptr(O2_required, __ argument_address(constant(0), noreg, 0));
1247 jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1248 }
1249 break;
1251 case _invokestatic_mh:
1252 case _invokespecial_mh:
1253 {
1254 __ load_heap_oop(G3_mh_vmtarget, G5_method); // target is a methodOop
1255 // Same as TemplateTable::invokestatic or invokespecial,
1256 // minus the CP setup and profiling:
1257 if (ek == _invokespecial_mh) {
1258 // Must load & check the first argument before entering the target method.
1259 __ load_method_handle_vmslots(O0_argslot, G3_method_handle, O1_scratch);
1260 __ ld_ptr(__ argument_address(O0_argslot, O0_argslot, -1), G3_method_handle);
1261 __ null_check(G3_method_handle);
1262 __ verify_oop(G3_method_handle);
1263 }
1264 jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1265 }
1266 break;
1268 case _invokevirtual_mh:
1269 {
1270 // Same as TemplateTable::invokevirtual,
1271 // minus the CP setup and profiling:
1273 // Pick out the vtable index and receiver offset from the MH,
1274 // and then we can discard it:
1275 Register O2_index = O2_scratch;
1276 __ load_method_handle_vmslots(O0_argslot, G3_method_handle, O1_scratch);
1277 __ ldsw(G3_dmh_vmindex, O2_index);
1278 // Note: The verifier allows us to ignore G3_mh_vmtarget.
1279 __ ld_ptr(__ argument_address(O0_argslot, O0_argslot, -1), G3_method_handle);
1280 __ null_check(G3_method_handle, oopDesc::klass_offset_in_bytes());
1282 // Get receiver klass:
1283 Register O0_klass = O0_argslot;
1284 __ load_klass(G3_method_handle, O0_klass);
1285 __ verify_oop(O0_klass);
1287 // Get target methodOop & entry point:
1288 const int base = instanceKlass::vtable_start_offset() * wordSize;
1289 assert(vtableEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
1291 __ sll_ptr(O2_index, LogBytesPerWord, O2_index);
1292 __ add(O0_klass, O2_index, O0_klass);
1293 Address vtable_entry_addr(O0_klass, base + vtableEntry::method_offset_in_bytes());
1294 __ ld_ptr(vtable_entry_addr, G5_method);
1296 jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1297 }
1298 break;
1300 case _invokeinterface_mh:
1301 {
1302 // Same as TemplateTable::invokeinterface,
1303 // minus the CP setup and profiling:
1304 __ load_method_handle_vmslots(O0_argslot, G3_method_handle, O1_scratch);
1305 Register O1_intf = O1_scratch;
1306 Register G5_index = G5_scratch;
1307 __ load_heap_oop(G3_mh_vmtarget, O1_intf);
1308 __ ldsw(G3_dmh_vmindex, G5_index);
1309 __ ld_ptr(__ argument_address(O0_argslot, O0_argslot, -1), G3_method_handle);
1310 __ null_check(G3_method_handle, oopDesc::klass_offset_in_bytes());
1312 // Get receiver klass:
1313 Register O0_klass = O0_argslot;
1314 __ load_klass(G3_method_handle, O0_klass);
1315 __ verify_oop(O0_klass);
1317 // Get interface:
1318 Label no_such_interface;
1319 __ verify_oop(O1_intf);
1320 __ lookup_interface_method(O0_klass, O1_intf,
1321 // Note: next two args must be the same:
1322 G5_index, G5_method,
1323 O2_scratch,
1324 O3_scratch,
1325 no_such_interface);
1327 jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1329 __ bind(no_such_interface);
1330 // Throw an exception.
1331 // For historical reasons, it will be IncompatibleClassChangeError.
1332 __ unimplemented("not tested yet");
1333 __ ld_ptr(Address(O1_intf, java_mirror_offset), O2_required); // required interface
1334 __ mov( O0_klass, O1_actual); // bad receiver
1335 __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
1336 __ delayed()->mov(Bytecodes::_invokeinterface, O0_code); // who is complaining?
1337 }
1338 break;
1340 case _bound_ref_mh:
1341 case _bound_int_mh:
1342 case _bound_long_mh:
1343 case _bound_ref_direct_mh:
1344 case _bound_int_direct_mh:
1345 case _bound_long_direct_mh:
1346 {
1347 const bool direct_to_method = (ek >= _bound_ref_direct_mh);
1348 BasicType arg_type = ek_bound_mh_arg_type(ek);
1349 int arg_slots = type2size[arg_type];
1351 // Make room for the new argument:
1352 load_vmargslot(_masm, G3_bmh_vmargslot, O0_argslot);
1353 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1355 insert_arg_slots(_masm, arg_slots * stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1357 // Store bound argument into the new stack slot:
1358 __ load_heap_oop(G3_bmh_argument, O1_scratch);
1359 if (arg_type == T_OBJECT) {
1360 __ st_ptr(O1_scratch, Address(O0_argslot, 0));
1361 } else {
1362 Address prim_value_addr(O1_scratch, java_lang_boxing_object::value_offset_in_bytes(arg_type));
1363 move_typed_arg(_masm, arg_type, false,
1364 prim_value_addr,
1365 Address(O0_argslot, 0),
1366 O2_scratch); // must be an even register for !_LP64 long moves (uses O2/O3)
1367 }
1369 if (direct_to_method) {
1370 __ load_heap_oop(G3_mh_vmtarget, G5_method); // target is a methodOop
1371 jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1372 } else {
1373 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle); // target is a methodOop
1374 __ verify_oop(G3_method_handle);
1375 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1376 }
1377 }
1378 break;
1380 case _adapter_opt_profiling:
1381 if (java_lang_invoke_CountingMethodHandle::vmcount_offset_in_bytes() != 0) {
1382 Address G3_mh_vmcount(G3_method_handle, java_lang_invoke_CountingMethodHandle::vmcount_offset_in_bytes());
1383 __ ld(G3_mh_vmcount, O1_scratch);
1384 __ add(O1_scratch, 1, O1_scratch);
1385 __ st(O1_scratch, G3_mh_vmcount);
1386 }
1387 // fall through
1389 case _adapter_retype_only:
1390 case _adapter_retype_raw:
1391 // Immediately jump to the next MH layer:
1392 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1393 __ verify_oop(G3_method_handle);
1394 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1395 // This is OK when all parameter types widen.
1396 // It is also OK when a return type narrows.
1397 break;
1399 case _adapter_check_cast:
1400 {
1401 // Check a reference argument before jumping to the next layer of MH:
1402 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1403 Address vmarg = __ argument_address(O0_argslot, O0_argslot);
1405 // What class are we casting to?
1406 Register O1_klass = O1_scratch; // Interesting AMH data.
1407 __ load_heap_oop(G3_amh_argument, O1_klass); // This is a Class object!
1408 load_klass_from_Class(_masm, O1_klass, O2_scratch, O3_scratch);
1410 Label L_done;
1411 __ ld_ptr(vmarg, O2_scratch);
1412 __ br_null_short(O2_scratch, Assembler::pn, L_done); // No cast if null.
1413 __ load_klass(O2_scratch, O2_scratch);
1415 // Live at this point:
1416 // - O0_argslot : argslot index in vmarg; may be required in the failing path
1417 // - O1_klass : klass required by the target method
1418 // - O2_scratch : argument klass to test
1419 // - G3_method_handle: adapter method handle
1420 __ check_klass_subtype(O2_scratch, O1_klass, O3_scratch, O4_scratch, L_done);
1422 // If we get here, the type check failed!
1423 __ load_heap_oop(G3_amh_argument, O2_required); // required class
1424 __ ld_ptr( vmarg, O1_actual); // bad object
1425 __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
1426 __ delayed()->mov(Bytecodes::_checkcast, O0_code); // who is complaining?
1428 __ BIND(L_done);
1429 // Get the new MH:
1430 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1431 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1432 }
1433 break;
1435 case _adapter_prim_to_prim:
1436 case _adapter_ref_to_prim:
1437 // Handled completely by optimized cases.
1438 __ stop("init_AdapterMethodHandle should not issue this");
1439 break;
1441 case _adapter_opt_i2i: // optimized subcase of adapt_prim_to_prim
1442 //case _adapter_opt_f2i: // optimized subcase of adapt_prim_to_prim
1443 case _adapter_opt_l2i: // optimized subcase of adapt_prim_to_prim
1444 case _adapter_opt_unboxi: // optimized subcase of adapt_ref_to_prim
1445 {
1446 // Perform an in-place conversion to int or an int subword.
1447 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1448 Address value;
1449 Address vmarg;
1450 bool value_left_justified = false;
1452 switch (ek) {
1453 case _adapter_opt_i2i:
1454 value = vmarg = __ argument_address(O0_argslot, O0_argslot);
1455 break;
1456 case _adapter_opt_l2i:
1457 {
1458 // just delete the extra slot
1459 #ifdef _LP64
1460 // In V9, longs are given 2 64-bit slots in the interpreter, but the
1461 // data is passed in only 1 slot.
1462 // Keep the second slot.
1463 __ add(__ argument_address(O0_argslot, O0_argslot, -1), O0_argslot);
1464 remove_arg_slots(_masm, -stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1465 value = Address(O0_argslot, 4); // Get least-significant 32-bit of 64-bit value.
1466 vmarg = Address(O0_argslot, Interpreter::stackElementSize);
1467 #else
1468 // Keep the first slot.
1469 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1470 remove_arg_slots(_masm, -stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1471 value = Address(O0_argslot, 0);
1472 vmarg = value;
1473 #endif
1474 }
1475 break;
1476 case _adapter_opt_unboxi:
1477 {
1478 vmarg = __ argument_address(O0_argslot, O0_argslot);
1479 // Load the value up from the heap.
1480 __ ld_ptr(vmarg, O1_scratch);
1481 int value_offset = java_lang_boxing_object::value_offset_in_bytes(T_INT);
1482 #ifdef ASSERT
1483 for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
1484 if (is_subword_type(BasicType(bt)))
1485 assert(value_offset == java_lang_boxing_object::value_offset_in_bytes(BasicType(bt)), "");
1486 }
1487 #endif
1488 __ null_check(O1_scratch, value_offset);
1489 value = Address(O1_scratch, value_offset);
1490 #ifdef _BIG_ENDIAN
1491 // Values stored in objects are packed.
1492 value_left_justified = true;
1493 #endif
1494 }
1495 break;
1496 default:
1497 ShouldNotReachHere();
1498 }
1500 // This check is required on _BIG_ENDIAN
1501 Register G5_vminfo = G5_scratch;
1502 __ ldsw(G3_amh_conversion, G5_vminfo);
1503 assert(CONV_VMINFO_SHIFT == 0, "preshifted");
1505 // Original 32-bit vmdata word must be of this form:
1506 // | MBZ:6 | signBitCount:8 | srcDstTypes:8 | conversionOp:8 |
1507 __ lduw(value, O1_scratch);
1508 if (!value_left_justified)
1509 __ sll(O1_scratch, G5_vminfo, O1_scratch);
1510 Label zero_extend, done;
1511 __ btst(CONV_VMINFO_SIGN_FLAG, G5_vminfo);
1512 __ br(Assembler::zero, false, Assembler::pn, zero_extend);
1513 __ delayed()->nop();
1515 // this path is taken for int->byte, int->short
1516 __ sra(O1_scratch, G5_vminfo, O1_scratch);
1517 __ ba_short(done);
1519 __ bind(zero_extend);
1520 // this is taken for int->char
1521 __ srl(O1_scratch, G5_vminfo, O1_scratch);
1523 __ bind(done);
1524 __ st(O1_scratch, vmarg);
1526 // Get the new MH:
1527 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1528 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1529 }
1530 break;
1532 case _adapter_opt_i2l: // optimized subcase of adapt_prim_to_prim
1533 case _adapter_opt_unboxl: // optimized subcase of adapt_ref_to_prim
1534 {
1535 // Perform an in-place int-to-long or ref-to-long conversion.
1536 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1538 // On big-endian machine we duplicate the slot and store the MSW
1539 // in the first slot.
1540 __ add(__ argument_address(O0_argslot, O0_argslot, 1), O0_argslot);
1542 insert_arg_slots(_masm, stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1544 Address arg_lsw(O0_argslot, 0);
1545 Address arg_msw(O0_argslot, -Interpreter::stackElementSize);
1547 switch (ek) {
1548 case _adapter_opt_i2l:
1549 {
1550 #ifdef _LP64
1551 __ ldsw(arg_lsw, O2_scratch); // Load LSW sign-extended
1552 #else
1553 __ ldsw(arg_lsw, O3_scratch); // Load LSW sign-extended
1554 __ srlx(O3_scratch, BitsPerInt, O2_scratch); // Move MSW value to lower 32-bits for std
1555 #endif
1556 __ st_long(O2_scratch, arg_msw); // Uses O2/O3 on !_LP64
1557 }
1558 break;
1559 case _adapter_opt_unboxl:
1560 {
1561 // Load the value up from the heap.
1562 __ ld_ptr(arg_lsw, O1_scratch);
1563 int value_offset = java_lang_boxing_object::value_offset_in_bytes(T_LONG);
1564 assert(value_offset == java_lang_boxing_object::value_offset_in_bytes(T_DOUBLE), "");
1565 __ null_check(O1_scratch, value_offset);
1566 __ ld_long(Address(O1_scratch, value_offset), O2_scratch); // Uses O2/O3 on !_LP64
1567 __ st_long(O2_scratch, arg_msw);
1568 }
1569 break;
1570 default:
1571 ShouldNotReachHere();
1572 }
1574 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1575 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1576 }
1577 break;
1579 case _adapter_opt_f2d: // optimized subcase of adapt_prim_to_prim
1580 case _adapter_opt_d2f: // optimized subcase of adapt_prim_to_prim
1581 {
1582 // perform an in-place floating primitive conversion
1583 __ unimplemented(entry_name(ek));
1584 }
1585 break;
1587 case _adapter_prim_to_ref:
1588 __ unimplemented(entry_name(ek)); // %%% FIXME: NYI
1589 break;
1591 case _adapter_swap_args:
1592 case _adapter_rot_args:
1593 // handled completely by optimized cases
1594 __ stop("init_AdapterMethodHandle should not issue this");
1595 break;
1597 case _adapter_opt_swap_1:
1598 case _adapter_opt_swap_2:
1599 case _adapter_opt_rot_1_up:
1600 case _adapter_opt_rot_1_down:
1601 case _adapter_opt_rot_2_up:
1602 case _adapter_opt_rot_2_down:
1603 {
1604 int swap_slots = ek_adapter_opt_swap_slots(ek);
1605 int rotate = ek_adapter_opt_swap_mode(ek);
1607 // 'argslot' is the position of the first argument to swap.
1608 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1609 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1610 if (VerifyMethodHandles)
1611 verify_argslot(_masm, O0_argslot, O2_scratch, "swap point must fall within current frame");
1613 // 'vminfo' is the second.
1614 Register O1_destslot = O1_scratch;
1615 load_conversion_vminfo(_masm, G3_amh_conversion, O1_destslot);
1616 __ add(__ argument_address(O1_destslot, O1_destslot), O1_destslot);
1617 if (VerifyMethodHandles)
1618 verify_argslot(_masm, O1_destslot, O2_scratch, "swap point must fall within current frame");
1620 assert(Interpreter::stackElementSize == wordSize, "else rethink use of wordSize here");
1621 if (!rotate) {
1622 // simple swap
1623 for (int i = 0; i < swap_slots; i++) {
1624 __ ld_ptr( Address(O0_argslot, i * wordSize), O2_scratch);
1625 __ ld_ptr( Address(O1_destslot, i * wordSize), O3_scratch);
1626 __ st_ptr(O3_scratch, Address(O0_argslot, i * wordSize));
1627 __ st_ptr(O2_scratch, Address(O1_destslot, i * wordSize));
1628 }
1629 } else {
1630 // A rotate is actually pair of moves, with an "odd slot" (or pair)
1631 // changing place with a series of other slots.
1632 // First, push the "odd slot", which is going to get overwritten
1633 switch (swap_slots) {
1634 case 2 : __ ld_ptr(Address(O0_argslot, 1 * wordSize), O4_scratch); // fall-thru
1635 case 1 : __ ld_ptr(Address(O0_argslot, 0 * wordSize), O3_scratch); break;
1636 default: ShouldNotReachHere();
1637 }
1638 if (rotate > 0) {
1639 // Here is rotate > 0:
1640 // (low mem) (high mem)
1641 // | dest: more_slots... | arg: odd_slot :arg+1 |
1642 // =>
1643 // | dest: odd_slot | dest+1: more_slots... :arg+1 |
1644 // work argslot down to destslot, copying contiguous data upwards
1645 // pseudo-code:
1646 // argslot = src_addr - swap_bytes
1647 // destslot = dest_addr
1648 // while (argslot >= destslot) *(argslot + swap_bytes) = *(argslot + 0), argslot--;
1649 move_arg_slots_up(_masm,
1650 O1_destslot,
1651 Address(O0_argslot, 0),
1652 swap_slots,
1653 O0_argslot, O2_scratch);
1654 } else {
1655 // Here is the other direction, rotate < 0:
1656 // (low mem) (high mem)
1657 // | arg: odd_slot | arg+1: more_slots... :dest+1 |
1658 // =>
1659 // | arg: more_slots... | dest: odd_slot :dest+1 |
1660 // work argslot up to destslot, copying contiguous data downwards
1661 // pseudo-code:
1662 // argslot = src_addr + swap_bytes
1663 // destslot = dest_addr
1664 // while (argslot <= destslot) *(argslot - swap_bytes) = *(argslot + 0), argslot++;
1665 // dest_slot denotes an exclusive upper limit
1666 int limit_bias = OP_ROT_ARGS_DOWN_LIMIT_BIAS;
1667 if (limit_bias != 0)
1668 __ add(O1_destslot, - limit_bias * wordSize, O1_destslot);
1669 move_arg_slots_down(_masm,
1670 Address(O0_argslot, swap_slots * wordSize),
1671 O1_destslot,
1672 -swap_slots,
1673 O0_argslot, O2_scratch);
1675 __ sub(O1_destslot, swap_slots * wordSize, O1_destslot);
1676 }
1677 // pop the original first chunk into the destination slot, now free
1678 switch (swap_slots) {
1679 case 2 : __ st_ptr(O4_scratch, Address(O1_destslot, 1 * wordSize)); // fall-thru
1680 case 1 : __ st_ptr(O3_scratch, Address(O1_destslot, 0 * wordSize)); break;
1681 default: ShouldNotReachHere();
1682 }
1683 }
1685 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1686 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1687 }
1688 break;
1690 case _adapter_dup_args:
1691 {
1692 // 'argslot' is the position of the first argument to duplicate.
1693 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1694 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1696 // 'stack_move' is negative number of words to duplicate.
1697 Register O1_stack_move = O1_scratch;
1698 load_stack_move(_masm, G3_amh_conversion, O1_stack_move);
1700 if (VerifyMethodHandles) {
1701 verify_argslots(_masm, O1_stack_move, O0_argslot, O2_scratch, O3_scratch, true,
1702 "copied argument(s) must fall within current frame");
1703 }
1705 // insert location is always the bottom of the argument list:
1706 __ neg(O1_stack_move);
1707 push_arg_slots(_masm, O0_argslot, O1_stack_move, O2_scratch, O3_scratch);
1709 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1710 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1711 }
1712 break;
1714 case _adapter_drop_args:
1715 {
1716 // 'argslot' is the position of the first argument to nuke.
1717 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1718 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1720 // 'stack_move' is number of words to drop.
1721 Register O1_stack_move = O1_scratch;
1722 load_stack_move(_masm, G3_amh_conversion, O1_stack_move);
1724 remove_arg_slots(_masm, O1_stack_move, O0_argslot, O2_scratch, O3_scratch, O4_scratch);
1726 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1727 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1728 }
1729 break;
1731 case _adapter_collect_args:
1732 case _adapter_fold_args:
1733 case _adapter_spread_args:
1734 // Handled completely by optimized cases.
1735 __ stop("init_AdapterMethodHandle should not issue this");
1736 break;
1738 case _adapter_opt_collect_ref:
1739 case _adapter_opt_collect_int:
1740 case _adapter_opt_collect_long:
1741 case _adapter_opt_collect_float:
1742 case _adapter_opt_collect_double:
1743 case _adapter_opt_collect_void:
1744 case _adapter_opt_collect_0_ref:
1745 case _adapter_opt_collect_1_ref:
1746 case _adapter_opt_collect_2_ref:
1747 case _adapter_opt_collect_3_ref:
1748 case _adapter_opt_collect_4_ref:
1749 case _adapter_opt_collect_5_ref:
1750 case _adapter_opt_filter_S0_ref:
1751 case _adapter_opt_filter_S1_ref:
1752 case _adapter_opt_filter_S2_ref:
1753 case _adapter_opt_filter_S3_ref:
1754 case _adapter_opt_filter_S4_ref:
1755 case _adapter_opt_filter_S5_ref:
1756 case _adapter_opt_collect_2_S0_ref:
1757 case _adapter_opt_collect_2_S1_ref:
1758 case _adapter_opt_collect_2_S2_ref:
1759 case _adapter_opt_collect_2_S3_ref:
1760 case _adapter_opt_collect_2_S4_ref:
1761 case _adapter_opt_collect_2_S5_ref:
1762 case _adapter_opt_fold_ref:
1763 case _adapter_opt_fold_int:
1764 case _adapter_opt_fold_long:
1765 case _adapter_opt_fold_float:
1766 case _adapter_opt_fold_double:
1767 case _adapter_opt_fold_void:
1768 case _adapter_opt_fold_1_ref:
1769 case _adapter_opt_fold_2_ref:
1770 case _adapter_opt_fold_3_ref:
1771 case _adapter_opt_fold_4_ref:
1772 case _adapter_opt_fold_5_ref:
1773 {
1774 // Given a fresh incoming stack frame, build a new ricochet frame.
1775 // On entry, TOS points at a return PC, and FP is the callers frame ptr.
1776 // RSI/R13 has the caller's exact stack pointer, which we must also preserve.
1777 // RCX contains an AdapterMethodHandle of the indicated kind.
1779 // Relevant AMH fields:
1780 // amh.vmargslot:
1781 // points to the trailing edge of the arguments
1782 // to filter, collect, or fold. For a boxing operation,
1783 // it points just after the single primitive value.
1784 // amh.argument:
1785 // recursively called MH, on |collect| arguments
1786 // amh.vmtarget:
1787 // final destination MH, on return value, etc.
1788 // amh.conversion.dest:
1789 // tells what is the type of the return value
1790 // (not needed here, since dest is also derived from ek)
1791 // amh.conversion.vminfo:
1792 // points to the trailing edge of the return value
1793 // when the vmtarget is to be called; this is
1794 // equal to vmargslot + (retained ? |collect| : 0)
1796 // Pass 0 or more argument slots to the recursive target.
1797 int collect_count_constant = ek_adapter_opt_collect_count(ek);
1799 // The collected arguments are copied from the saved argument list:
1800 int collect_slot_constant = ek_adapter_opt_collect_slot(ek);
1802 assert(ek_orig == _adapter_collect_args ||
1803 ek_orig == _adapter_fold_args, "");
1804 bool retain_original_args = (ek_orig == _adapter_fold_args);
1806 // The return value is replaced (or inserted) at the 'vminfo' argslot.
1807 // Sometimes we can compute this statically.
1808 int dest_slot_constant = -1;
1809 if (!retain_original_args)
1810 dest_slot_constant = collect_slot_constant;
1811 else if (collect_slot_constant >= 0 && collect_count_constant >= 0)
1812 // We are preserving all the arguments, and the return value is prepended,
1813 // so the return slot is to the left (above) the |collect| sequence.
1814 dest_slot_constant = collect_slot_constant + collect_count_constant;
1816 // Replace all those slots by the result of the recursive call.
1817 // The result type can be one of ref, int, long, float, double, void.
1818 // In the case of void, nothing is pushed on the stack after return.
1819 BasicType dest = ek_adapter_opt_collect_type(ek);
1820 assert(dest == type2wfield[dest], "dest is a stack slot type");
1821 int dest_count = type2size[dest];
1822 assert(dest_count == 1 || dest_count == 2 || (dest_count == 0 && dest == T_VOID), "dest has a size");
1824 // Choose a return continuation.
1825 EntryKind ek_ret = _adapter_opt_return_any;
1826 if (dest != T_CONFLICT && OptimizeMethodHandles) {
1827 switch (dest) {
1828 case T_INT : ek_ret = _adapter_opt_return_int; break;
1829 case T_LONG : ek_ret = _adapter_opt_return_long; break;
1830 case T_FLOAT : ek_ret = _adapter_opt_return_float; break;
1831 case T_DOUBLE : ek_ret = _adapter_opt_return_double; break;
1832 case T_OBJECT : ek_ret = _adapter_opt_return_ref; break;
1833 case T_VOID : ek_ret = _adapter_opt_return_void; break;
1834 default : ShouldNotReachHere();
1835 }
1836 if (dest == T_OBJECT && dest_slot_constant >= 0) {
1837 EntryKind ek_try = EntryKind(_adapter_opt_return_S0_ref + dest_slot_constant);
1838 if (ek_try <= _adapter_opt_return_LAST &&
1839 ek_adapter_opt_return_slot(ek_try) == dest_slot_constant) {
1840 ek_ret = ek_try;
1841 }
1842 }
1843 assert(ek_adapter_opt_return_type(ek_ret) == dest, "");
1844 }
1846 // Already pushed: ... keep1 | collect | keep2 |
1848 // Push a few extra argument words, if we need them to store the return value.
1849 {
1850 int extra_slots = 0;
1851 if (retain_original_args) {
1852 extra_slots = dest_count;
1853 } else if (collect_count_constant == -1) {
1854 extra_slots = dest_count; // collect_count might be zero; be generous
1855 } else if (dest_count > collect_count_constant) {
1856 extra_slots = (dest_count - collect_count_constant);
1857 } else {
1858 // else we know we have enough dead space in |collect| to repurpose for return values
1859 }
1860 if (extra_slots != 0) {
1861 __ sub(SP, round_to(extra_slots, 2) * Interpreter::stackElementSize, SP);
1862 }
1863 }
1865 // Set up Ricochet Frame.
1866 __ mov(SP, O5_savedSP); // record SP for the callee
1868 // One extra (empty) slot for outgoing target MH (see Gargs computation below).
1869 __ save_frame(2); // Note: we need to add 2 slots since frame::memory_parameter_word_sp_offset is 23.
1871 // Note: Gargs is live throughout the following, until we make our recursive call.
1872 // And the RF saves a copy in L4_saved_args_base.
1874 RicochetFrame::enter_ricochet_frame(_masm, G3_method_handle, Gargs,
1875 entry(ek_ret)->from_interpreted_entry());
1877 // Compute argument base:
1878 // Set up Gargs for current frame, extra (empty) slot is for outgoing target MH (space reserved by save_frame above).
1879 __ add(FP, STACK_BIAS - (1 * Interpreter::stackElementSize), Gargs);
1881 // Now pushed: ... keep1 | collect | keep2 | extra | [RF]
1883 #ifdef ASSERT
1884 if (VerifyMethodHandles && dest != T_CONFLICT) {
1885 BLOCK_COMMENT("verify AMH.conv.dest {");
1886 extract_conversion_dest_type(_masm, RicochetFrame::L5_conversion, O1_scratch);
1887 Label L_dest_ok;
1888 __ cmp(O1_scratch, (int) dest);
1889 __ br(Assembler::equal, false, Assembler::pt, L_dest_ok);
1890 __ delayed()->nop();
1891 if (dest == T_INT) {
1892 for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
1893 if (is_subword_type(BasicType(bt))) {
1894 __ cmp(O1_scratch, (int) bt);
1895 __ br(Assembler::equal, false, Assembler::pt, L_dest_ok);
1896 __ delayed()->nop();
1897 }
1898 }
1899 }
1900 __ stop("bad dest in AMH.conv");
1901 __ BIND(L_dest_ok);
1902 BLOCK_COMMENT("} verify AMH.conv.dest");
1903 }
1904 #endif //ASSERT
1906 // Find out where the original copy of the recursive argument sequence begins.
1907 Register O0_coll = O0_scratch;
1908 {
1909 RegisterOrConstant collect_slot = collect_slot_constant;
1910 if (collect_slot_constant == -1) {
1911 load_vmargslot(_masm, G3_amh_vmargslot, O1_scratch);
1912 collect_slot = O1_scratch;
1913 }
1914 // collect_slot might be 0, but we need the move anyway.
1915 __ add(RicochetFrame::L4_saved_args_base, __ argument_offset(collect_slot, collect_slot.register_or_noreg()), O0_coll);
1916 // O0_coll now points at the trailing edge of |collect| and leading edge of |keep2|
1917 }
1919 // Replace the old AMH with the recursive MH. (No going back now.)
1920 // In the case of a boxing call, the recursive call is to a 'boxer' method,
1921 // such as Integer.valueOf or Long.valueOf. In the case of a filter
1922 // or collect call, it will take one or more arguments, transform them,
1923 // and return some result, to store back into argument_base[vminfo].
1924 __ load_heap_oop(G3_amh_argument, G3_method_handle);
1925 if (VerifyMethodHandles) verify_method_handle(_masm, G3_method_handle, O1_scratch, O2_scratch);
1927 // Calculate |collect|, the number of arguments we are collecting.
1928 Register O1_collect_count = O1_scratch;
1929 RegisterOrConstant collect_count;
1930 if (collect_count_constant < 0) {
1931 __ load_method_handle_vmslots(O1_collect_count, G3_method_handle, O2_scratch);
1932 collect_count = O1_collect_count;
1933 } else {
1934 collect_count = collect_count_constant;
1935 #ifdef ASSERT
1936 if (VerifyMethodHandles) {
1937 BLOCK_COMMENT("verify collect_count_constant {");
1938 __ load_method_handle_vmslots(O3_scratch, G3_method_handle, O2_scratch);
1939 Label L_count_ok;
1940 __ cmp_and_br_short(O3_scratch, collect_count_constant, Assembler::equal, Assembler::pt, L_count_ok);
1941 __ stop("bad vminfo in AMH.conv");
1942 __ BIND(L_count_ok);
1943 BLOCK_COMMENT("} verify collect_count_constant");
1944 }
1945 #endif //ASSERT
1946 }
1948 // copy |collect| slots directly to TOS:
1949 push_arg_slots(_masm, O0_coll, collect_count, O2_scratch, O3_scratch);
1950 // Now pushed: ... keep1 | collect | keep2 | RF... | collect |
1951 // O0_coll still points at the trailing edge of |collect| and leading edge of |keep2|
1953 // If necessary, adjust the saved arguments to make room for the eventual return value.
1954 // Normal adjustment: ... keep1 | +dest+ | -collect- | keep2 | RF... | collect |
1955 // If retaining args: ... keep1 | +dest+ | collect | keep2 | RF... | collect |
1956 // In the non-retaining case, this might move keep2 either up or down.
1957 // We don't have to copy the whole | RF... collect | complex,
1958 // but we must adjust RF.saved_args_base.
1959 // Also, from now on, we will forget about the original copy of |collect|.
1960 // If we are retaining it, we will treat it as part of |keep2|.
1961 // For clarity we will define |keep3| = |collect|keep2| or |keep2|.
1963 BLOCK_COMMENT("adjust trailing arguments {");
1964 // Compare the sizes of |+dest+| and |-collect-|, which are opposed opening and closing movements.
1965 int open_count = dest_count;
1966 RegisterOrConstant close_count = collect_count_constant;
1967 Register O1_close_count = O1_collect_count;
1968 if (retain_original_args) {
1969 close_count = constant(0);
1970 } else if (collect_count_constant == -1) {
1971 close_count = O1_collect_count;
1972 }
1974 // How many slots need moving? This is simply dest_slot (0 => no |keep3|).
1975 RegisterOrConstant keep3_count;
1976 Register O2_keep3_count = O2_scratch;
1977 if (dest_slot_constant < 0) {
1978 extract_conversion_vminfo(_masm, RicochetFrame::L5_conversion, O2_keep3_count);
1979 keep3_count = O2_keep3_count;
1980 } else {
1981 keep3_count = dest_slot_constant;
1982 #ifdef ASSERT
1983 if (VerifyMethodHandles && dest_slot_constant < 0) {
1984 BLOCK_COMMENT("verify dest_slot_constant {");
1985 extract_conversion_vminfo(_masm, RicochetFrame::L5_conversion, O3_scratch);
1986 Label L_vminfo_ok;
1987 __ cmp_and_br_short(O3_scratch, dest_slot_constant, Assembler::equal, Assembler::pt, L_vminfo_ok);
1988 __ stop("bad vminfo in AMH.conv");
1989 __ BIND(L_vminfo_ok);
1990 BLOCK_COMMENT("} verify dest_slot_constant");
1991 }
1992 #endif //ASSERT
1993 }
1995 // tasks remaining:
1996 bool move_keep3 = (!keep3_count.is_constant() || keep3_count.as_constant() != 0);
1997 bool stomp_dest = (NOT_DEBUG(dest == T_OBJECT) DEBUG_ONLY(dest_count != 0));
1998 bool fix_arg_base = (!close_count.is_constant() || open_count != close_count.as_constant());
2000 // Old and new argument locations (based at slot 0).
2001 // Net shift (&new_argv - &old_argv) is (close_count - open_count).
2002 bool zero_open_count = (open_count == 0); // remember this bit of info
2003 if (move_keep3 && fix_arg_base) {
2004 // It will be easier to have everything in one register:
2005 if (close_count.is_register()) {
2006 // Deduct open_count from close_count register to get a clean +/- value.
2007 __ sub(close_count.as_register(), open_count, close_count.as_register());
2008 } else {
2009 close_count = close_count.as_constant() - open_count;
2010 }
2011 open_count = 0;
2012 }
2013 Register L4_old_argv = RicochetFrame::L4_saved_args_base;
2014 Register O3_new_argv = O3_scratch;
2015 if (fix_arg_base) {
2016 __ add(L4_old_argv, __ argument_offset(close_count, O4_scratch), O3_new_argv,
2017 -(open_count * Interpreter::stackElementSize));
2018 }
2020 // First decide if any actual data are to be moved.
2021 // We can skip if (a) |keep3| is empty, or (b) the argument list size didn't change.
2022 // (As it happens, all movements involve an argument list size change.)
2024 // If there are variable parameters, use dynamic checks to skip around the whole mess.
2025 Label L_done;
2026 if (keep3_count.is_register()) {
2027 __ cmp_and_br_short(keep3_count.as_register(), 0, Assembler::equal, Assembler::pn, L_done);
2028 }
2029 if (close_count.is_register()) {
2030 __ cmp_and_br_short(close_count.as_register(), open_count, Assembler::equal, Assembler::pn, L_done);
2031 }
2033 if (move_keep3 && fix_arg_base) {
2034 bool emit_move_down = false, emit_move_up = false, emit_guard = false;
2035 if (!close_count.is_constant()) {
2036 emit_move_down = emit_guard = !zero_open_count;
2037 emit_move_up = true;
2038 } else if (open_count != close_count.as_constant()) {
2039 emit_move_down = (open_count > close_count.as_constant());
2040 emit_move_up = !emit_move_down;
2041 }
2042 Label L_move_up;
2043 if (emit_guard) {
2044 __ cmp(close_count.as_register(), open_count);
2045 __ br(Assembler::greater, false, Assembler::pn, L_move_up);
2046 __ delayed()->nop();
2047 }
2049 if (emit_move_down) {
2050 // Move arguments down if |+dest+| > |-collect-|
2051 // (This is rare, except when arguments are retained.)
2052 // This opens space for the return value.
2053 if (keep3_count.is_constant()) {
2054 for (int i = 0; i < keep3_count.as_constant(); i++) {
2055 __ ld_ptr( Address(L4_old_argv, i * Interpreter::stackElementSize), O4_scratch);
2056 __ st_ptr(O4_scratch, Address(O3_new_argv, i * Interpreter::stackElementSize) );
2057 }
2058 } else {
2059 // Live: O1_close_count, O2_keep3_count, O3_new_argv
2060 Register argv_top = O0_scratch;
2061 __ add(L4_old_argv, __ argument_offset(keep3_count, O4_scratch), argv_top);
2062 move_arg_slots_down(_masm,
2063 Address(L4_old_argv, 0), // beginning of old argv
2064 argv_top, // end of old argv
2065 close_count, // distance to move down (must be negative)
2066 O4_scratch, G5_scratch);
2067 }
2068 }
2070 if (emit_guard) {
2071 __ ba_short(L_done); // assumes emit_move_up is true also
2072 __ BIND(L_move_up);
2073 }
2075 if (emit_move_up) {
2076 // Move arguments up if |+dest+| < |-collect-|
2077 // (This is usual, except when |keep3| is empty.)
2078 // This closes up the space occupied by the now-deleted collect values.
2079 if (keep3_count.is_constant()) {
2080 for (int i = keep3_count.as_constant() - 1; i >= 0; i--) {
2081 __ ld_ptr( Address(L4_old_argv, i * Interpreter::stackElementSize), O4_scratch);
2082 __ st_ptr(O4_scratch, Address(O3_new_argv, i * Interpreter::stackElementSize) );
2083 }
2084 } else {
2085 Address argv_top(L4_old_argv, __ argument_offset(keep3_count, O4_scratch));
2086 // Live: O1_close_count, O2_keep3_count, O3_new_argv
2087 move_arg_slots_up(_masm,
2088 L4_old_argv, // beginning of old argv
2089 argv_top, // end of old argv
2090 close_count, // distance to move up (must be positive)
2091 O4_scratch, G5_scratch);
2092 }
2093 }
2094 }
2095 __ BIND(L_done);
2097 if (fix_arg_base) {
2098 // adjust RF.saved_args_base
2099 __ mov(O3_new_argv, RicochetFrame::L4_saved_args_base);
2100 }
2102 if (stomp_dest) {
2103 // Stomp the return slot, so it doesn't hold garbage.
2104 // This isn't strictly necessary, but it may help detect bugs.
2105 __ set(RicochetFrame::RETURN_VALUE_PLACEHOLDER, O4_scratch);
2106 __ st_ptr(O4_scratch, Address(RicochetFrame::L4_saved_args_base,
2107 __ argument_offset(keep3_count, keep3_count.register_or_noreg()))); // uses O2_keep3_count
2108 }
2109 BLOCK_COMMENT("} adjust trailing arguments");
2111 BLOCK_COMMENT("do_recursive_call");
2112 __ mov(SP, O5_savedSP); // record SP for the callee
2113 __ set(ExternalAddress(SharedRuntime::ricochet_blob()->bounce_addr() - frame::pc_return_offset), O7);
2114 // The globally unique bounce address has two purposes:
2115 // 1. It helps the JVM recognize this frame (frame::is_ricochet_frame).
2116 // 2. When returned to, it cuts back the stack and redirects control flow
2117 // to the return handler.
2118 // The return handler will further cut back the stack when it takes
2119 // down the RF. Perhaps there is a way to streamline this further.
2121 // State during recursive call:
2122 // ... keep1 | dest | dest=42 | keep3 | RF... | collect | bounce_pc |
2123 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
2124 }
2125 break;
2127 case _adapter_opt_return_ref:
2128 case _adapter_opt_return_int:
2129 case _adapter_opt_return_long:
2130 case _adapter_opt_return_float:
2131 case _adapter_opt_return_double:
2132 case _adapter_opt_return_void:
2133 case _adapter_opt_return_S0_ref:
2134 case _adapter_opt_return_S1_ref:
2135 case _adapter_opt_return_S2_ref:
2136 case _adapter_opt_return_S3_ref:
2137 case _adapter_opt_return_S4_ref:
2138 case _adapter_opt_return_S5_ref:
2139 {
2140 BasicType dest_type_constant = ek_adapter_opt_return_type(ek);
2141 int dest_slot_constant = ek_adapter_opt_return_slot(ek);
2143 if (VerifyMethodHandles) RicochetFrame::verify_clean(_masm);
2145 if (dest_slot_constant == -1) {
2146 // The current stub is a general handler for this dest_type.
2147 // It can be called from _adapter_opt_return_any below.
2148 // Stash the address in a little table.
2149 assert((dest_type_constant & CONV_TYPE_MASK) == dest_type_constant, "oob");
2150 address return_handler = __ pc();
2151 _adapter_return_handlers[dest_type_constant] = return_handler;
2152 if (dest_type_constant == T_INT) {
2153 // do the subword types too
2154 for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
2155 if (is_subword_type(BasicType(bt)) &&
2156 _adapter_return_handlers[bt] == NULL) {
2157 _adapter_return_handlers[bt] = return_handler;
2158 }
2159 }
2160 }
2161 }
2163 // On entry to this continuation handler, make Gargs live again.
2164 __ mov(RicochetFrame::L4_saved_args_base, Gargs);
2166 Register O7_temp = O7;
2167 Register O5_vminfo = O5;
2169 RegisterOrConstant dest_slot = dest_slot_constant;
2170 if (dest_slot_constant == -1) {
2171 extract_conversion_vminfo(_masm, RicochetFrame::L5_conversion, O5_vminfo);
2172 dest_slot = O5_vminfo;
2173 }
2174 // Store the result back into the argslot.
2175 // This code uses the interpreter calling sequence, in which the return value
2176 // is usually left in the TOS register, as defined by InterpreterMacroAssembler::pop.
2177 // There are certain irregularities with floating point values, which can be seen
2178 // in TemplateInterpreterGenerator::generate_return_entry_for.
2179 move_return_value(_masm, dest_type_constant, __ argument_address(dest_slot, O7_temp));
2181 RicochetFrame::leave_ricochet_frame(_masm, G3_method_handle, I5_savedSP, I7);
2183 // Load the final target and go.
2184 if (VerifyMethodHandles) verify_method_handle(_masm, G3_method_handle, O0_scratch, O1_scratch);
2185 __ restore(I5_savedSP, G0, SP);
2186 __ jump_to_method_handle_entry(G3_method_handle, O0_scratch);
2187 __ illtrap(0);
2188 }
2189 break;
2191 case _adapter_opt_return_any:
2192 {
2193 Register O7_temp = O7;
2194 Register O5_dest_type = O5;
2196 if (VerifyMethodHandles) RicochetFrame::verify_clean(_masm);
2197 extract_conversion_dest_type(_masm, RicochetFrame::L5_conversion, O5_dest_type);
2198 __ set(ExternalAddress((address) &_adapter_return_handlers[0]), O7_temp);
2199 __ sll_ptr(O5_dest_type, LogBytesPerWord, O5_dest_type);
2200 __ ld_ptr(O7_temp, O5_dest_type, O7_temp);
2202 #ifdef ASSERT
2203 { Label L_ok;
2204 __ br_notnull_short(O7_temp, Assembler::pt, L_ok);
2205 __ stop("bad method handle return");
2206 __ BIND(L_ok);
2207 }
2208 #endif //ASSERT
2209 __ JMP(O7_temp, 0);
2210 __ delayed()->nop();
2211 }
2212 break;
2214 case _adapter_opt_spread_0:
2215 case _adapter_opt_spread_1_ref:
2216 case _adapter_opt_spread_2_ref:
2217 case _adapter_opt_spread_3_ref:
2218 case _adapter_opt_spread_4_ref:
2219 case _adapter_opt_spread_5_ref:
2220 case _adapter_opt_spread_ref:
2221 case _adapter_opt_spread_byte:
2222 case _adapter_opt_spread_char:
2223 case _adapter_opt_spread_short:
2224 case _adapter_opt_spread_int:
2225 case _adapter_opt_spread_long:
2226 case _adapter_opt_spread_float:
2227 case _adapter_opt_spread_double:
2228 {
2229 // spread an array out into a group of arguments
2230 int length_constant = ek_adapter_opt_spread_count(ek);
2231 bool length_can_be_zero = (length_constant == 0);
2232 if (length_constant < 0) {
2233 // some adapters with variable length must handle the zero case
2234 if (!OptimizeMethodHandles ||
2235 ek_adapter_opt_spread_type(ek) != T_OBJECT)
2236 length_can_be_zero = true;
2237 }
2239 // find the address of the array argument
2240 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
2241 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
2243 // O0_argslot points both to the array and to the first output arg
2244 Address vmarg = Address(O0_argslot, 0);
2246 // Get the array value.
2247 Register O1_array = O1_scratch;
2248 Register O2_array_klass = O2_scratch;
2249 BasicType elem_type = ek_adapter_opt_spread_type(ek);
2250 int elem_slots = type2size[elem_type]; // 1 or 2
2251 int array_slots = 1; // array is always a T_OBJECT
2252 int length_offset = arrayOopDesc::length_offset_in_bytes();
2253 int elem0_offset = arrayOopDesc::base_offset_in_bytes(elem_type);
2254 __ ld_ptr(vmarg, O1_array);
2256 Label L_array_is_empty, L_insert_arg_space, L_copy_args, L_args_done;
2257 if (length_can_be_zero) {
2258 // handle the null pointer case, if zero is allowed
2259 Label L_skip;
2260 if (length_constant < 0) {
2261 load_conversion_vminfo(_masm, G3_amh_conversion, O3_scratch);
2262 __ cmp_zero_and_br(Assembler::notZero, O3_scratch, L_skip);
2263 __ delayed()->nop(); // to avoid back-to-back cbcond instructions
2264 }
2265 __ br_null_short(O1_array, Assembler::pn, L_array_is_empty);
2266 __ BIND(L_skip);
2267 }
2268 __ null_check(O1_array, oopDesc::klass_offset_in_bytes());
2269 __ load_klass(O1_array, O2_array_klass);
2271 // Check the array type.
2272 Register O3_klass = O3_scratch;
2273 __ load_heap_oop(G3_amh_argument, O3_klass); // this is a Class object!
2274 load_klass_from_Class(_masm, O3_klass, O4_scratch, G5_scratch);
2276 Label L_ok_array_klass, L_bad_array_klass, L_bad_array_length;
2277 __ check_klass_subtype(O2_array_klass, O3_klass, O4_scratch, G5_scratch, L_ok_array_klass);
2278 // If we get here, the type check failed!
2279 __ ba_short(L_bad_array_klass);
2280 __ BIND(L_ok_array_klass);
2282 // Check length.
2283 if (length_constant >= 0) {
2284 __ ldsw(Address(O1_array, length_offset), O4_scratch);
2285 __ cmp(O4_scratch, length_constant);
2286 } else {
2287 Register O3_vminfo = O3_scratch;
2288 load_conversion_vminfo(_masm, G3_amh_conversion, O3_vminfo);
2289 __ ldsw(Address(O1_array, length_offset), O4_scratch);
2290 __ cmp(O3_vminfo, O4_scratch);
2291 }
2292 __ br(Assembler::notEqual, false, Assembler::pn, L_bad_array_length);
2293 __ delayed()->nop();
2295 Register O2_argslot_limit = O2_scratch;
2297 // Array length checks out. Now insert any required stack slots.
2298 if (length_constant == -1) {
2299 // Form a pointer to the end of the affected region.
2300 __ add(O0_argslot, Interpreter::stackElementSize, O2_argslot_limit);
2301 // 'stack_move' is negative number of words to insert
2302 // This number already accounts for elem_slots.
2303 Register O3_stack_move = O3_scratch;
2304 load_stack_move(_masm, G3_amh_conversion, O3_stack_move);
2305 __ cmp(O3_stack_move, 0);
2306 assert(stack_move_unit() < 0, "else change this comparison");
2307 __ br(Assembler::less, false, Assembler::pn, L_insert_arg_space);
2308 __ delayed()->nop();
2309 __ br(Assembler::equal, false, Assembler::pn, L_copy_args);
2310 __ delayed()->nop();
2311 // single argument case, with no array movement
2312 __ BIND(L_array_is_empty);
2313 remove_arg_slots(_masm, -stack_move_unit() * array_slots,
2314 O0_argslot, O1_scratch, O2_scratch, O3_scratch);
2315 __ ba_short(L_args_done); // no spreading to do
2316 __ BIND(L_insert_arg_space);
2317 // come here in the usual case, stack_move < 0 (2 or more spread arguments)
2318 // Live: O1_array, O2_argslot_limit, O3_stack_move
2319 insert_arg_slots(_masm, O3_stack_move,
2320 O0_argslot, O4_scratch, G5_scratch, O1_scratch);
2321 // reload from rdx_argslot_limit since rax_argslot is now decremented
2322 __ ld_ptr(Address(O2_argslot_limit, -Interpreter::stackElementSize), O1_array);
2323 } else if (length_constant >= 1) {
2324 int new_slots = (length_constant * elem_slots) - array_slots;
2325 insert_arg_slots(_masm, new_slots * stack_move_unit(),
2326 O0_argslot, O2_scratch, O3_scratch, O4_scratch);
2327 } else if (length_constant == 0) {
2328 __ BIND(L_array_is_empty);
2329 remove_arg_slots(_masm, -stack_move_unit() * array_slots,
2330 O0_argslot, O1_scratch, O2_scratch, O3_scratch);
2331 } else {
2332 ShouldNotReachHere();
2333 }
2335 // Copy from the array to the new slots.
2336 // Note: Stack change code preserves integrity of O0_argslot pointer.
2337 // So even after slot insertions, O0_argslot still points to first argument.
2338 // Beware: Arguments that are shallow on the stack are deep in the array,
2339 // and vice versa. So a downward-growing stack (the usual) has to be copied
2340 // elementwise in reverse order from the source array.
2341 __ BIND(L_copy_args);
2342 if (length_constant == -1) {
2343 // [O0_argslot, O2_argslot_limit) is the area we are inserting into.
2344 // Array element [0] goes at O0_argslot_limit[-wordSize].
2345 Register O1_source = O1_array;
2346 __ add(Address(O1_array, elem0_offset), O1_source);
2347 Register O4_fill_ptr = O4_scratch;
2348 __ mov(O2_argslot_limit, O4_fill_ptr);
2349 Label L_loop;
2350 __ BIND(L_loop);
2351 __ add(O4_fill_ptr, -Interpreter::stackElementSize * elem_slots, O4_fill_ptr);
2352 move_typed_arg(_masm, elem_type, true,
2353 Address(O1_source, 0), Address(O4_fill_ptr, 0),
2354 O2_scratch); // must be an even register for !_LP64 long moves (uses O2/O3)
2355 __ add(O1_source, type2aelembytes(elem_type), O1_source);
2356 __ cmp_and_brx_short(O4_fill_ptr, O0_argslot, Assembler::greaterUnsigned, Assembler::pt, L_loop);
2357 } else if (length_constant == 0) {
2358 // nothing to copy
2359 } else {
2360 int elem_offset = elem0_offset;
2361 int slot_offset = length_constant * Interpreter::stackElementSize;
2362 for (int index = 0; index < length_constant; index++) {
2363 slot_offset -= Interpreter::stackElementSize * elem_slots; // fill backward
2364 move_typed_arg(_masm, elem_type, true,
2365 Address(O1_array, elem_offset), Address(O0_argslot, slot_offset),
2366 O2_scratch); // must be an even register for !_LP64 long moves (uses O2/O3)
2367 elem_offset += type2aelembytes(elem_type);
2368 }
2369 }
2370 __ BIND(L_args_done);
2372 // Arguments are spread. Move to next method handle.
2373 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
2374 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
2376 __ BIND(L_bad_array_klass);
2377 assert(!vmarg.uses(O2_required), "must be different registers");
2378 __ load_heap_oop(Address(O2_array_klass, java_mirror_offset), O2_required); // required class
2379 __ ld_ptr( vmarg, O1_actual); // bad object
2380 __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
2381 __ delayed()->mov(Bytecodes::_aaload, O0_code); // who is complaining?
2383 __ bind(L_bad_array_length);
2384 assert(!vmarg.uses(O2_required), "must be different registers");
2385 __ mov( G3_method_handle, O2_required); // required class
2386 __ ld_ptr(vmarg, O1_actual); // bad object
2387 __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
2388 __ delayed()->mov(Bytecodes::_arraylength, O0_code); // who is complaining?
2389 }
2390 break;
2392 default:
2393 DEBUG_ONLY(tty->print_cr("bad ek=%d (%s)", (int)ek, entry_name(ek)));
2394 ShouldNotReachHere();
2395 }
2396 BLOCK_COMMENT(err_msg("} Entry %s", entry_name(ek)));
2398 address me_cookie = MethodHandleEntry::start_compiled_entry(_masm, interp_entry);
2399 __ unimplemented(entry_name(ek)); // %%% FIXME: NYI
2401 init_entry(ek, MethodHandleEntry::finish_compiled_entry(_masm, me_cookie));
2402 }