1 /*
2 * Copyright (c) 1999, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "c1/c1_MacroAssembler.hpp"
27 #include "c1/c1_Runtime1.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/barrierSetAssembler.hpp"
31 #include "gc/shared/collectedHeap.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "oops/arrayOop.hpp"
34 #include "oops/markOop.hpp"
35 #include "runtime/basicLock.hpp"
36 #include "runtime/biasedLocking.hpp"
37 #include "runtime/frame.inline.hpp"
38 #include "runtime/os.hpp"
39 #include "runtime/sharedRuntime.hpp"
40 #include "runtime/stubRoutines.hpp"
41
42 int C1_MacroAssembler::lock_object(Register hdr, Register obj, Register disp_hdr, Register scratch, Label& slow_case) {
43 const int aligned_mask = BytesPerWord -1;
44 const int hdr_offset = oopDesc::mark_offset_in_bytes();
45 assert(hdr == rax, "hdr must be rax, for the cmpxchg instruction");
46 assert(hdr != obj && hdr != disp_hdr && obj != disp_hdr, "registers must be different");
47 Label done;
48 int null_check_offset = -1;
49
50 verify_oop(obj);
51
52 // save object being locked into the BasicObjectLock
53 movptr(Address(disp_hdr, BasicObjectLock::obj_offset_in_bytes()), obj);
54
55 if (UseBiasedLocking) {
56 assert(scratch != noreg, "should have scratch register at this point");
57 null_check_offset = biased_locking_enter(disp_hdr, obj, hdr, scratch, false, done, &slow_case);
58 } else {
59 null_check_offset = offset();
60 }
61
62 // Load object header
63 movptr(hdr, Address(obj, hdr_offset));
64 // and mark it as unlocked
65 orptr(hdr, markOopDesc::unlocked_value);
66 if (EnableValhalla && !UseBiasedLocking) {
67 // Mask always_locked bit such that we go to the slow path if object is a value type
68 andptr(hdr, ~markOopDesc::biased_lock_bit_in_place);
69 }
70 // save unlocked object header into the displaced header location on the stack
71 movptr(Address(disp_hdr, 0), hdr);
72 // test if object header is still the same (i.e. unlocked), and if so, store the
73 // displaced header address in the object header - if it is not the same, get the
74 // object header instead
75 MacroAssembler::lock(); // must be immediately before cmpxchg!
76 cmpxchgptr(disp_hdr, Address(obj, hdr_offset));
77 // if the object header was the same, we're done
78 if (PrintBiasedLockingStatistics) {
79 cond_inc32(Assembler::equal,
80 ExternalAddress((address)BiasedLocking::fast_path_entry_count_addr()));
81 }
82 jcc(Assembler::equal, done);
83 // if the object header was not the same, it is now in the hdr register
84 // => test if it is a stack pointer into the same stack (recursive locking), i.e.:
85 //
86 // 1) (hdr & aligned_mask) == 0
87 // 2) rsp <= hdr
88 // 3) hdr <= rsp + page_size
89 //
90 // these 3 tests can be done by evaluating the following expression:
91 //
92 // (hdr - rsp) & (aligned_mask - page_size)
93 //
94 // assuming both the stack pointer and page_size have their least
95 // significant 2 bits cleared and page_size is a power of 2
96 subptr(hdr, rsp);
97 andptr(hdr, aligned_mask - os::vm_page_size());
98 // for recursive locking, the result is zero => save it in the displaced header
99 // location (NULL in the displaced hdr location indicates recursive locking)
100 movptr(Address(disp_hdr, 0), hdr);
101 // otherwise we don't care about the result and handle locking via runtime call
102 jcc(Assembler::notZero, slow_case);
103 // done
104 bind(done);
105 return null_check_offset;
106 }
107
108
109 void C1_MacroAssembler::unlock_object(Register hdr, Register obj, Register disp_hdr, Label& slow_case) {
110 const int aligned_mask = BytesPerWord -1;
111 const int hdr_offset = oopDesc::mark_offset_in_bytes();
112 assert(disp_hdr == rax, "disp_hdr must be rax, for the cmpxchg instruction");
113 assert(hdr != obj && hdr != disp_hdr && obj != disp_hdr, "registers must be different");
114 Label done;
115
116 if (UseBiasedLocking) {
117 // load object
118 movptr(obj, Address(disp_hdr, BasicObjectLock::obj_offset_in_bytes()));
119 biased_locking_exit(obj, hdr, done);
120 }
121
122 // load displaced header
123 movptr(hdr, Address(disp_hdr, 0));
124 // if the loaded hdr is NULL we had recursive locking
125 testptr(hdr, hdr);
126 // if we had recursive locking, we are done
127 jcc(Assembler::zero, done);
128 if (!UseBiasedLocking) {
129 // load object
130 movptr(obj, Address(disp_hdr, BasicObjectLock::obj_offset_in_bytes()));
131 }
132 verify_oop(obj);
133 // test if object header is pointing to the displaced header, and if so, restore
134 // the displaced header in the object - if the object header is not pointing to
135 // the displaced header, get the object header instead
136 MacroAssembler::lock(); // must be immediately before cmpxchg!
137 cmpxchgptr(hdr, Address(obj, hdr_offset));
138 // if the object header was not pointing to the displaced header,
139 // we do unlocking via runtime call
140 jcc(Assembler::notEqual, slow_case);
141 // done
142 bind(done);
143 }
144
145
146 // Defines obj, preserves var_size_in_bytes
147 void C1_MacroAssembler::try_allocate(Register obj, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2, Label& slow_case) {
148 if (UseTLAB) {
149 tlab_allocate(noreg, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
150 } else {
151 eden_allocate(noreg, obj, var_size_in_bytes, con_size_in_bytes, t1, slow_case);
152 }
153 }
154
155
156 void C1_MacroAssembler::initialize_header(Register obj, Register klass, Register len, Register t1, Register t2) {
157 assert_different_registers(obj, klass, len);
158 if ((UseBiasedLocking || EnableValhalla) && !len->is_valid()) {
159 // Need to copy markOopDesc::always_locked_pattern for values.
160 assert_different_registers(obj, klass, len, t1, t2);
161 movptr(t1, Address(klass, Klass::prototype_header_offset()));
162 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), t1);
163 } else {
164 // This assumes that all prototype bits fit in an int32_t
165 movptr(Address(obj, oopDesc::mark_offset_in_bytes ()), (int32_t)(intptr_t)markOopDesc::prototype());
166 }
167 #ifdef _LP64
168 if (UseCompressedClassPointers) { // Take care not to kill klass
169 movptr(t1, klass);
170 encode_klass_not_null(t1);
171 movl(Address(obj, oopDesc::klass_offset_in_bytes()), t1);
172 } else
173 #endif
174 {
175 movptr(Address(obj, oopDesc::klass_offset_in_bytes()), klass);
176 }
177
178 if (len->is_valid()) {
179 movl(Address(obj, arrayOopDesc::length_offset_in_bytes()), len);
180 }
181 #ifdef _LP64
182 else if (UseCompressedClassPointers) {
183 xorptr(t1, t1);
184 store_klass_gap(obj, t1);
185 }
186 #endif
187 }
188
189
190 // preserves obj, destroys len_in_bytes
191 void C1_MacroAssembler::initialize_body(Register obj, Register len_in_bytes, int hdr_size_in_bytes, Register t1) {
192 assert(hdr_size_in_bytes >= 0, "header size must be positive or 0");
193 Label done;
194
195 // len_in_bytes is positive and ptr sized
196 subptr(len_in_bytes, hdr_size_in_bytes);
197 jcc(Assembler::zero, done);
198 zero_memory(obj, len_in_bytes, hdr_size_in_bytes, t1);
199 bind(done);
200 }
201
202
203 void C1_MacroAssembler::allocate_object(Register obj, Register t1, Register t2, int header_size, int object_size, Register klass, Label& slow_case) {
204 assert(obj == rax, "obj must be in rax, for cmpxchg");
205 assert_different_registers(obj, t1, t2); // XXX really?
206 assert(header_size >= 0 && object_size >= header_size, "illegal sizes");
207
208 try_allocate(obj, noreg, object_size * BytesPerWord, t1, t2, slow_case);
209
210 initialize_object(obj, klass, noreg, object_size * HeapWordSize, t1, t2, UseTLAB);
211 }
212
213 void C1_MacroAssembler::initialize_object(Register obj, Register klass, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2, bool is_tlab_allocated) {
214 assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0,
215 "con_size_in_bytes is not multiple of alignment");
216 const int hdr_size_in_bytes = instanceOopDesc::header_size() * HeapWordSize;
217
218 initialize_header(obj, klass, noreg, t1, t2);
219
220 if (!(UseTLAB && ZeroTLAB && is_tlab_allocated)) {
221 // clear rest of allocated space
222 const Register t1_zero = t1;
223 const Register index = t2;
224 const int threshold = 6 * BytesPerWord; // approximate break even point for code size (see comments below)
225 if (var_size_in_bytes != noreg) {
226 mov(index, var_size_in_bytes);
227 initialize_body(obj, index, hdr_size_in_bytes, t1_zero);
228 } else if (con_size_in_bytes <= threshold) {
229 // use explicit null stores
230 // code size = 2 + 3*n bytes (n = number of fields to clear)
231 xorptr(t1_zero, t1_zero); // use t1_zero reg to clear memory (shorter code)
232 for (int i = hdr_size_in_bytes; i < con_size_in_bytes; i += BytesPerWord)
233 movptr(Address(obj, i), t1_zero);
234 } else if (con_size_in_bytes > hdr_size_in_bytes) {
235 // use loop to null out the fields
236 // code size = 16 bytes for even n (n = number of fields to clear)
237 // initialize last object field first if odd number of fields
238 xorptr(t1_zero, t1_zero); // use t1_zero reg to clear memory (shorter code)
239 movptr(index, (con_size_in_bytes - hdr_size_in_bytes) >> 3);
240 // initialize last object field if constant size is odd
241 if (((con_size_in_bytes - hdr_size_in_bytes) & 4) != 0)
242 movptr(Address(obj, con_size_in_bytes - (1*BytesPerWord)), t1_zero);
243 // initialize remaining object fields: rdx is a multiple of 2
244 { Label loop;
245 bind(loop);
246 movptr(Address(obj, index, Address::times_8, hdr_size_in_bytes - (1*BytesPerWord)),
247 t1_zero);
248 NOT_LP64(movptr(Address(obj, index, Address::times_8, hdr_size_in_bytes - (2*BytesPerWord)),
249 t1_zero);)
250 decrement(index);
251 jcc(Assembler::notZero, loop);
252 }
253 }
254 }
255
256 if (CURRENT_ENV->dtrace_alloc_probes()) {
257 assert(obj == rax, "must be");
258 call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id)));
259 }
260
261 verify_oop(obj);
262 }
263
264 void C1_MacroAssembler::allocate_array(Register obj, Register len, Register t1, Register t2, int header_size, Address::ScaleFactor f, Register klass, Label& slow_case) {
265 assert(obj == rax, "obj must be in rax, for cmpxchg");
266 assert_different_registers(obj, len, t1, t2, klass);
267
268 // determine alignment mask
269 assert(!(BytesPerWord & 1), "must be a multiple of 2 for masking code to work");
270
271 // check for negative or excessive length
272 cmpptr(len, (int32_t)max_array_allocation_length);
273 jcc(Assembler::above, slow_case);
274
275 const Register arr_size = t2; // okay to be the same
276 // align object end
277 movptr(arr_size, (int32_t)header_size * BytesPerWord + MinObjAlignmentInBytesMask);
278 lea(arr_size, Address(arr_size, len, f));
279 andptr(arr_size, ~MinObjAlignmentInBytesMask);
280
281 try_allocate(obj, arr_size, 0, t1, t2, slow_case);
282
283 initialize_header(obj, klass, len, t1, t2);
284
285 // clear rest of allocated space
286 const Register len_zero = len;
287 initialize_body(obj, arr_size, header_size * BytesPerWord, len_zero);
288
289 if (CURRENT_ENV->dtrace_alloc_probes()) {
290 assert(obj == rax, "must be");
291 call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id)));
292 }
293
294 verify_oop(obj);
295 }
296
297
298
299 void C1_MacroAssembler::inline_cache_check(Register receiver, Register iCache) {
300 verify_oop(receiver);
301 // explicit NULL check not needed since load from [klass_offset] causes a trap
302 // check against inline cache
303 assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()), "must add explicit null check");
304 int start_offset = offset();
305
306 if (UseCompressedClassPointers) {
307 load_klass(rscratch1, receiver);
308 cmpptr(rscratch1, iCache);
309 } else {
310 cmpptr(iCache, Address(receiver, oopDesc::klass_offset_in_bytes()));
311 }
312 // if icache check fails, then jump to runtime routine
313 // Note: RECEIVER must still contain the receiver!
314 jump_cc(Assembler::notEqual,
315 RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
316 const int ic_cmp_size = LP64_ONLY(10) NOT_LP64(9);
317 assert(UseCompressedClassPointers || offset() - start_offset == ic_cmp_size, "check alignment in emit_method_entry");
318 }
319
320
321 void C1_MacroAssembler::build_frame(int frame_size_in_bytes, int bang_size_in_bytes, bool needs_stack_repair, Label* verified_value_entry_label) {
322 assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
323 // Make sure there is enough stack space for this method's activation.
324 // Note that we do this before doing an enter(). This matches the
325 // ordering of C2's stack overflow check / rsp decrement and allows
326 // the SharedRuntime stack overflow handling to be consistent
327 // between the two compilers.
328 generate_stack_overflow_check(bang_size_in_bytes);
329
330 if (!needs_stack_repair && verified_value_entry_label != NULL) {
331 bind(*verified_value_entry_label);
332 }
333 push(rbp);
334 if (PreserveFramePointer) {
335 mov(rbp, rsp);
336 }
337 #ifdef TIERED
338 // c2 leaves fpu stack dirty. Clean it on entry
339 if (UseSSE < 2 ) {
340 empty_FPU_stack();
341 }
342 #endif // TIERED
343 decrement(rsp, frame_size_in_bytes); // does not emit code for frame_size == 0
344 if (needs_stack_repair) {
345 movptr(Address(rsp, frame_size_in_bytes - wordSize), frame_size_in_bytes
346 + wordSize // skip over pushed rbp
347 + wordSize); // skip over RA pushed by caller
348 if (verified_value_entry_label != NULL) {
349 bind(*verified_value_entry_label);
350 }
351 }
352
353 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
354 bs->nmethod_entry_barrier(this);
355 }
356
357
358 void C1_MacroAssembler::remove_frame(int frame_size_in_bytes, bool needs_stack_repair) {
359 if (!needs_stack_repair) {
360 increment(rsp, frame_size_in_bytes); // Does not emit code for frame_size == 0
361 pop(rbp);
362 } else {
363 movq(r13, Address(rsp, frame_size_in_bytes + wordSize)); // return address
364 movq(rbp, Address(rsp, frame_size_in_bytes));
365 addq(rsp, Address(rsp, frame_size_in_bytes - wordSize)); // now we are back to caller frame, without the outgoing returned address
366 push(r13); // restore the returned address, as pushed by caller
367 }
368 }
369
370
371 void C1_MacroAssembler::verified_value_entry() {
372 if (C1Breakpoint || VerifyFPU || !UseStackBanging) {
373 // Verified Entry first instruction should be 5 bytes long for correct
374 // patching by patch_verified_entry().
375 //
376 // C1Breakpoint and VerifyFPU have one byte first instruction.
377 // Also first instruction will be one byte "push(rbp)" if stack banging
378 // code is not generated (see build_frame() above).
379 // For all these cases generate long instruction first.
380 fat_nop();
381 }
382 if (C1Breakpoint)int3();
383 // build frame
384 verify_FPU(0, "method_entry");
385 }
386
387 int C1_MacroAssembler::scalarized_entry(const CompiledEntrySignature *ces, int frame_size_in_bytes, int bang_size_in_bytes, Label& verified_value_entry_label, bool is_value_ro_entry) {
388 if (C1Breakpoint || VerifyFPU || !UseStackBanging) {
389 // Verified Entry first instruction should be 5 bytes long for correct
390 // patching by patch_verified_entry().
391 //
392 // C1Breakpoint and VerifyFPU have one byte first instruction.
393 // Also first instruction will be one byte "push(rbp)" if stack banging
394 // code is not generated (see build_frame() above).
395 // For all these cases generate long instruction first.
396 fat_nop();
397 }
398 if (C1Breakpoint)int3();
399 verify_FPU(0, "method_entry");
400
401 // FIXME -- call runtime only if we cannot in-line allocate all the incoming value args.
402 push(rbp); // Create a temp frame so we can call into runtime // FIXME: need to be able to handle GC during the call
403 if (PreserveFramePointer) {
404 mov(rbp, rsp);
405 }
406 subptr(rsp, frame_size_in_bytes);
407 movptr(rbx, (intptr_t)(ces->method()));
408 if (is_value_ro_entry) {
409 call(RuntimeAddress(Runtime1::entry_for(Runtime1::buffer_value_args_no_receiver_id)));
410 } else {
411 call(RuntimeAddress(Runtime1::entry_for(Runtime1::buffer_value_args_id)));
412 }
413 int rt_call_offset = offset();
414
415 addptr(rsp, frame_size_in_bytes);
416 pop(rbp);
417
418 assert(ValueTypePassFieldsAsArgs, "sanity");
419
420 GrowableArray<SigEntry>* sig = &ces->sig();
421 GrowableArray<SigEntry>* sig_cc = is_value_ro_entry ? &ces->sig_cc_ro() : &ces->sig_cc();
422 VMRegPair* regs = ces->regs();
423 VMRegPair* regs_cc = is_value_ro_entry ? ces->regs_cc_ro() : ces->regs_cc();
424 int args_on_stack = ces->args_on_stack();
425 int args_on_stack_cc = is_value_ro_entry ? ces->args_on_stack_cc_ro() : ces->args_on_stack_cc();
426
427 assert(sig->length() <= sig_cc->length(), "Zero-sized value class not allowed!");
428 BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, sig_cc->length());
429 int args_passed = sig->length();
430 int args_passed_cc = SigEntry::fill_sig_bt(sig_cc, sig_bt);
431
432 int extra_stack_offset = wordSize; // tos is return address.
433 int sp_inc = shuffle_value_args(true, is_value_ro_entry, extra_stack_offset, sig_bt, sig_cc,
434 args_passed_cc, args_on_stack_cc, regs_cc, // from
435 args_passed, args_on_stack, regs); // to
436
437 if (sp_inc != 0) {
438 assert(sp_inc > 0, "stack should not shrink");
439 generate_stack_overflow_check(bang_size_in_bytes);
440 push(rbp);
441 if (PreserveFramePointer) {
442 mov(rbp, rsp);
443 }
444 #ifdef TIERED
445 // c2 leaves fpu stack dirty. Clean it on entry
446 if (UseSSE < 2 ) {
447 empty_FPU_stack();
448 }
449 #endif // TIERED
450 decrement(rsp, frame_size_in_bytes);
451 movptr(Address(rsp, frame_size_in_bytes - wordSize), frame_size_in_bytes +
452 + wordSize // pushed rbp
453 + wordSize // returned address pushed by the stack extension code
454 + sp_inc); // stack extension
455 }
456
457 jmp(verified_value_entry_label);
458 return rt_call_offset;
459 }
460
461 void C1_MacroAssembler::load_parameter(int offset_in_words, Register reg) {
462 // rbp, + 0: link
463 // + 1: return address
464 // + 2: argument with offset 0
465 // + 3: argument with offset 1
466 // + 4: ...
467
468 movptr(reg, Address(rbp, (offset_in_words + 2) * BytesPerWord));
469 }
470
471 #ifndef PRODUCT
472
473 void C1_MacroAssembler::verify_stack_oop(int stack_offset) {
474 if (!VerifyOops) return;
475 verify_oop_addr(Address(rsp, stack_offset));
476 }
477
478 void C1_MacroAssembler::verify_not_null_oop(Register r) {
479 if (!VerifyOops) return;
480 Label not_null;
481 testptr(r, r);
482 jcc(Assembler::notZero, not_null);
483 stop("non-null oop required");
484 bind(not_null);
485 verify_oop(r);
486 }
487
488 void C1_MacroAssembler::invalidate_registers(bool inv_rax, bool inv_rbx, bool inv_rcx, bool inv_rdx, bool inv_rsi, bool inv_rdi) {
489 #ifdef ASSERT
490 if (inv_rax) movptr(rax, 0xDEAD);
491 if (inv_rbx) movptr(rbx, 0xDEAD);
492 if (inv_rcx) movptr(rcx, 0xDEAD);
493 if (inv_rdx) movptr(rdx, 0xDEAD);
494 if (inv_rsi) movptr(rsi, 0xDEAD);
495 if (inv_rdi) movptr(rdi, 0xDEAD);
496 #endif
497 }
498
499 #endif // ifndef PRODUCT