1 /*
2 * Copyright (c) 1999, 2022, 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 "compiler/compilerDefinitions.inline.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/barrierSetAssembler.hpp"
31 #include "gc/shared/collectedHeap.hpp"
32 #include "gc/shared/tlab_globals.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "oops/arrayOop.hpp"
35 #include "oops/markWord.hpp"
36 #include "runtime/basicLock.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, 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 null_check_offset = offset();
56
57 if (DiagnoseSyncOnValueBasedClasses != 0) {
58 load_klass(hdr, obj, rscratch1);
59 movl(hdr, Address(hdr, Klass::access_flags_offset()));
60 testl(hdr, JVM_ACC_IS_VALUE_BASED_CLASS);
61 jcc(Assembler::notZero, slow_case);
62 }
63
64 // Load object header
65 movptr(hdr, Address(obj, hdr_offset));
66 // and mark it as unlocked
67 orptr(hdr, markWord::unlocked_value);
68 if (EnableValhalla) {
69 // Mask inline_type bit such that we go to the slow path if object is an inline type
70 andptr(hdr, ~((int) markWord::inline_type_bit_in_place));
71 }
72 // save unlocked object header into the displaced header location on the stack
73 movptr(Address(disp_hdr, 0), hdr);
74 // test if object header is still the same (i.e. unlocked), and if so, store the
75 // displaced header address in the object header - if it is not the same, get the
76 // object header instead
77 MacroAssembler::lock(); // must be immediately before cmpxchg!
78 cmpxchgptr(disp_hdr, Address(obj, hdr_offset));
79 // if the object header was the same, we're done
80 jcc(Assembler::equal, done);
81 // if the object header was not the same, it is now in the hdr register
82 // => test if it is a stack pointer into the same stack (recursive locking), i.e.:
83 //
84 // 1) (hdr & aligned_mask) == 0
85 // 2) rsp <= hdr
86 // 3) hdr <= rsp + page_size
87 //
88 // these 3 tests can be done by evaluating the following expression:
89 //
90 // (hdr - rsp) & (aligned_mask - page_size)
91 //
92 // assuming both the stack pointer and page_size have their least
93 // significant 2 bits cleared and page_size is a power of 2
94 subptr(hdr, rsp);
95 andptr(hdr, aligned_mask - os::vm_page_size());
96 // for recursive locking, the result is zero => save it in the displaced header
97 // location (NULL in the displaced hdr location indicates recursive locking)
98 movptr(Address(disp_hdr, 0), hdr);
99 // otherwise we don't care about the result and handle locking via runtime call
100 jcc(Assembler::notZero, slow_case);
101 // done
102 bind(done);
103
104 inc_held_monitor_count();
105
106 return null_check_offset;
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 // load displaced header
117 movptr(hdr, Address(disp_hdr, 0));
118 // if the loaded hdr is NULL we had recursive locking
119 testptr(hdr, hdr);
120 // if we had recursive locking, we are done
121 jcc(Assembler::zero, done);
122 // load object
123 movptr(obj, Address(disp_hdr, BasicObjectLock::obj_offset_in_bytes()));
124
125 verify_oop(obj);
126 // test if object header is pointing to the displaced header, and if so, restore
127 // the displaced header in the object - if the object header is not pointing to
128 // the displaced header, get the object header instead
129 MacroAssembler::lock(); // must be immediately before cmpxchg!
130 cmpxchgptr(hdr, Address(obj, hdr_offset));
131 // if the object header was not pointing to the displaced header,
132 // we do unlocking via runtime call
133 jcc(Assembler::notEqual, slow_case);
134 // done
135 bind(done);
136
137 dec_held_monitor_count();
138 }
139
140
141 // Defines obj, preserves var_size_in_bytes
142 void C1_MacroAssembler::try_allocate(Register obj, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2, Label& slow_case) {
143 if (UseTLAB) {
144 tlab_allocate(noreg, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
145 } else {
146 jmp(slow_case);
147 }
148 }
149
150
151 void C1_MacroAssembler::initialize_header(Register obj, Register klass, Register len, Register t1, Register t2) {
152 assert_different_registers(obj, klass, len);
153 if (EnableValhalla) {
154 // Need to copy markWord::prototype header for klass
155 assert_different_registers(obj, klass, len, t1, t2);
156 movptr(t1, Address(klass, Klass::prototype_header_offset()));
157 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), t1);
158 } else {
159 // This assumes that all prototype bits fit in an int32_t
160 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), checked_cast<int32_t>(markWord::prototype().value()));
161 }
162 #ifdef _LP64
163 if (UseCompressedClassPointers) { // Take care not to kill klass
164 movptr(t1, klass);
165 encode_klass_not_null(t1, rscratch1);
166 movl(Address(obj, oopDesc::klass_offset_in_bytes()), t1);
167 } else
168 #endif
169 {
170 movptr(Address(obj, oopDesc::klass_offset_in_bytes()), klass);
171 }
172
173 if (len->is_valid()) {
174 movl(Address(obj, arrayOopDesc::length_offset_in_bytes()), len);
175 }
176 #ifdef _LP64
177 else if (UseCompressedClassPointers) {
178 xorptr(t1, t1);
179 store_klass_gap(obj, t1);
180 }
181 #endif
182 }
183
184
185 // preserves obj, destroys len_in_bytes
186 void C1_MacroAssembler::initialize_body(Register obj, Register len_in_bytes, int hdr_size_in_bytes, Register t1) {
187 assert(hdr_size_in_bytes >= 0, "header size must be positive or 0");
188 Label done;
189
190 // len_in_bytes is positive and ptr sized
191 subptr(len_in_bytes, hdr_size_in_bytes);
192 jcc(Assembler::zero, done);
193 zero_memory(obj, len_in_bytes, hdr_size_in_bytes, t1);
194 bind(done);
195 }
196
197
198 void C1_MacroAssembler::allocate_object(Register obj, Register t1, Register t2, int header_size, int object_size, Register klass, Label& slow_case) {
199 assert(obj == rax, "obj must be in rax, for cmpxchg");
200 assert_different_registers(obj, t1, t2); // XXX really?
201 assert(header_size >= 0 && object_size >= header_size, "illegal sizes");
202
203 try_allocate(obj, noreg, object_size * BytesPerWord, t1, t2, slow_case);
204
205 initialize_object(obj, klass, noreg, object_size * HeapWordSize, t1, t2, UseTLAB);
206 }
207
208 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) {
209 assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0,
210 "con_size_in_bytes is not multiple of alignment");
211 const int hdr_size_in_bytes = instanceOopDesc::header_size() * HeapWordSize;
212
213 initialize_header(obj, klass, noreg, t1, t2);
214
215 if (!(UseTLAB && ZeroTLAB && is_tlab_allocated)) {
216 // clear rest of allocated space
217 const Register t1_zero = t1;
218 const Register index = t2;
219 const int threshold = 6 * BytesPerWord; // approximate break even point for code size (see comments below)
220 if (var_size_in_bytes != noreg) {
221 mov(index, var_size_in_bytes);
222 initialize_body(obj, index, hdr_size_in_bytes, t1_zero);
223 } else if (con_size_in_bytes <= threshold) {
224 // use explicit null stores
225 // code size = 2 + 3*n bytes (n = number of fields to clear)
226 xorptr(t1_zero, t1_zero); // use t1_zero reg to clear memory (shorter code)
227 for (int i = hdr_size_in_bytes; i < con_size_in_bytes; i += BytesPerWord)
228 movptr(Address(obj, i), t1_zero);
229 } else if (con_size_in_bytes > hdr_size_in_bytes) {
230 // use loop to null out the fields
231 // code size = 16 bytes for even n (n = number of fields to clear)
232 // initialize last object field first if odd number of fields
233 xorptr(t1_zero, t1_zero); // use t1_zero reg to clear memory (shorter code)
234 movptr(index, (con_size_in_bytes - hdr_size_in_bytes) >> 3);
235 // initialize last object field if constant size is odd
236 if (((con_size_in_bytes - hdr_size_in_bytes) & 4) != 0)
237 movptr(Address(obj, con_size_in_bytes - (1*BytesPerWord)), t1_zero);
238 // initialize remaining object fields: rdx is a multiple of 2
239 { Label loop;
240 bind(loop);
241 movptr(Address(obj, index, Address::times_8, hdr_size_in_bytes - (1*BytesPerWord)),
242 t1_zero);
243 NOT_LP64(movptr(Address(obj, index, Address::times_8, hdr_size_in_bytes - (2*BytesPerWord)),
244 t1_zero);)
245 decrement(index);
246 jcc(Assembler::notZero, loop);
247 }
248 }
249 }
250
251 if (CURRENT_ENV->dtrace_alloc_probes()) {
252 assert(obj == rax, "must be");
253 call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id)));
254 }
255
256 verify_oop(obj);
257 }
258
259 void C1_MacroAssembler::allocate_array(Register obj, Register len, Register t1, Register t2, int header_size, Address::ScaleFactor f, Register klass, Label& slow_case) {
260 assert(obj == rax, "obj must be in rax, for cmpxchg");
261 assert_different_registers(obj, len, t1, t2, klass);
262
263 // determine alignment mask
264 assert(!(BytesPerWord & 1), "must be a multiple of 2 for masking code to work");
265
266 // check for negative or excessive length
267 cmpptr(len, checked_cast<int32_t>(max_array_allocation_length));
268 jcc(Assembler::above, slow_case);
269
270 const Register arr_size = t2; // okay to be the same
271 // align object end
272 movptr(arr_size, header_size * BytesPerWord + MinObjAlignmentInBytesMask);
273 lea(arr_size, Address(arr_size, len, f));
274 andptr(arr_size, ~MinObjAlignmentInBytesMask);
275
276 try_allocate(obj, arr_size, 0, t1, t2, slow_case);
277
278 initialize_header(obj, klass, len, t1, t2);
279
280 // clear rest of allocated space
281 const Register len_zero = len;
282 initialize_body(obj, arr_size, header_size * BytesPerWord, len_zero);
283
284 if (CURRENT_ENV->dtrace_alloc_probes()) {
285 assert(obj == rax, "must be");
286 call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id)));
287 }
288
289 verify_oop(obj);
290 }
291
292
293
294 void C1_MacroAssembler::inline_cache_check(Register receiver, Register iCache) {
295 verify_oop(receiver);
296 // explicit NULL check not needed since load from [klass_offset] causes a trap
297 // check against inline cache
298 assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()), "must add explicit null check");
299 int start_offset = offset();
300
301 if (UseCompressedClassPointers) {
302 load_klass(rscratch1, receiver, rscratch2);
303 cmpptr(rscratch1, iCache);
304 } else {
305 cmpptr(iCache, Address(receiver, oopDesc::klass_offset_in_bytes()));
306 }
307 // if icache check fails, then jump to runtime routine
308 // Note: RECEIVER must still contain the receiver!
309 jump_cc(Assembler::notEqual,
310 RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
311 const int ic_cmp_size = LP64_ONLY(10) NOT_LP64(9);
312 assert(UseCompressedClassPointers || offset() - start_offset == ic_cmp_size, "check alignment in emit_method_entry");
313 }
314
315 void C1_MacroAssembler::build_frame_helper(int frame_size_in_bytes, int sp_offset_for_orig_pc, int sp_inc, bool has_scalarized_args, bool needs_stack_repair) {
316 push(rbp);
317 if (PreserveFramePointer) {
318 mov(rbp, rsp);
319 }
320 #if !defined(_LP64) && defined(COMPILER2)
321 if (UseSSE < 2 && !CompilerConfig::is_c1_only_no_jvmci()) {
322 // c2 leaves fpu stack dirty. Clean it on entry
323 empty_FPU_stack();
324 }
325 #endif // !_LP64 && COMPILER2
326 decrement(rsp, frame_size_in_bytes);
327
328 if (needs_stack_repair) {
329 // Save stack increment (also account for fixed framesize and rbp)
330 assert((sp_inc & (StackAlignmentInBytes-1)) == 0, "stack increment not aligned");
331 int real_frame_size = sp_inc + frame_size_in_bytes + wordSize;
332 movptr(Address(rsp, frame_size_in_bytes - wordSize), real_frame_size);
333 }
334 if (has_scalarized_args) {
335 // Initialize orig_pc to detect deoptimization during buffering in the entry points
336 movptr(Address(rsp, sp_offset_for_orig_pc), 0);
337 }
338 }
339
340 void C1_MacroAssembler::build_frame(int frame_size_in_bytes, int bang_size_in_bytes, int sp_offset_for_orig_pc, bool needs_stack_repair, bool has_scalarized_args, Label* verified_inline_entry_label) {
341 // Make sure there is enough stack space for this method's activation.
342 // Note that we do this before doing an enter(). This matches the
343 // ordering of C2's stack overflow check / rsp decrement and allows
344 // the SharedRuntime stack overflow handling to be consistent
345 // between the two compilers.
346 assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
347 generate_stack_overflow_check(bang_size_in_bytes);
348
349 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, 0, has_scalarized_args, needs_stack_repair);
350
351 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
352 // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
353 bs->nmethod_entry_barrier(this, NULL /* slow_path */, NULL /* continuation */);
354
355 if (verified_inline_entry_label != NULL) {
356 // Jump here from the scalarized entry points that already created the frame.
357 bind(*verified_inline_entry_label);
358 }
359 }
360
361 void C1_MacroAssembler::verified_entry(bool breakAtEntry) {
362 if (breakAtEntry || VerifyFPU) {
363 // Verified Entry first instruction should be 5 bytes long for correct
364 // patching by patch_verified_entry().
365 //
366 // Breakpoint and VerifyFPU have one byte first instruction.
367 // Also first instruction will be one byte "push(rbp)" if stack banging
368 // code is not generated (see build_frame() above).
369 // For all these cases generate long instruction first.
370 fat_nop();
371 }
372 if (breakAtEntry) int3();
373 // build frame
374 IA32_ONLY( verify_FPU(0, "method_entry"); )
375 }
376
377 int C1_MacroAssembler::scalarized_entry(const CompiledEntrySignature* ces, int frame_size_in_bytes, int bang_size_in_bytes, int sp_offset_for_orig_pc, Label& verified_inline_entry_label, bool is_inline_ro_entry) {
378 assert(InlineTypePassFieldsAsArgs, "sanity");
379 // Make sure there is enough stack space for this method's activation.
380 assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
381 generate_stack_overflow_check(bang_size_in_bytes);
382
383 GrowableArray<SigEntry>* sig = &ces->sig();
384 GrowableArray<SigEntry>* sig_cc = is_inline_ro_entry ? &ces->sig_cc_ro() : &ces->sig_cc();
385 VMRegPair* regs = ces->regs();
386 VMRegPair* regs_cc = is_inline_ro_entry ? ces->regs_cc_ro() : ces->regs_cc();
387 int args_on_stack = ces->args_on_stack();
388 int args_on_stack_cc = is_inline_ro_entry ? ces->args_on_stack_cc_ro() : ces->args_on_stack_cc();
389
390 assert(sig->length() <= sig_cc->length(), "Zero-sized inline class not allowed!");
391 BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, sig_cc->length());
392 int args_passed = sig->length();
393 int args_passed_cc = SigEntry::fill_sig_bt(sig_cc, sig_bt);
394
395 // Create a temp frame so we can call into the runtime. It must be properly set up to accommodate GC.
396 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, 0, true, ces->c1_needs_stack_repair());
397
398 // The runtime call might safepoint, make sure nmethod entry barrier is executed
399 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
400 // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
401 bs->nmethod_entry_barrier(this, NULL /* slow_path */, NULL /* continuation */);
402
403 // FIXME -- call runtime only if we cannot in-line allocate all the incoming inline type args.
404 movptr(rbx, (intptr_t)(ces->method()));
405 if (is_inline_ro_entry) {
406 call(RuntimeAddress(Runtime1::entry_for(Runtime1::buffer_inline_args_no_receiver_id)));
407 } else {
408 call(RuntimeAddress(Runtime1::entry_for(Runtime1::buffer_inline_args_id)));
409 }
410 int rt_call_offset = offset();
411
412 // Remove the temp frame
413 addptr(rsp, frame_size_in_bytes);
414 pop(rbp);
415
416 // Check if we need to extend the stack for packing
417 int sp_inc = 0;
418 if (args_on_stack > args_on_stack_cc) {
419 sp_inc = extend_stack_for_inline_args(args_on_stack);
420 }
421
422 shuffle_inline_args(true, is_inline_ro_entry, sig_cc,
423 args_passed_cc, args_on_stack_cc, regs_cc, // from
424 args_passed, args_on_stack, regs, // to
425 sp_inc, rax);
426
427 // Create the real frame. Below jump will then skip over the stack banging and frame
428 // setup code in the verified_inline_entry (which has a different real_frame_size).
429 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, sp_inc, true, ces->c1_needs_stack_repair());
430
431 jmp(verified_inline_entry_label);
432 return rt_call_offset;
433 }
434
435 void C1_MacroAssembler::load_parameter(int offset_in_words, Register reg) {
436 // rbp, + 0: link
437 // + 1: return address
438 // + 2: argument with offset 0
439 // + 3: argument with offset 1
440 // + 4: ...
441
442 movptr(reg, Address(rbp, (offset_in_words + 2) * BytesPerWord));
443 }
444
445 #ifndef PRODUCT
446
447 void C1_MacroAssembler::verify_stack_oop(int stack_offset) {
448 if (!VerifyOops) return;
449 verify_oop_addr(Address(rsp, stack_offset));
450 }
451
452 void C1_MacroAssembler::verify_not_null_oop(Register r) {
453 if (!VerifyOops) return;
454 Label not_null;
455 testptr(r, r);
456 jcc(Assembler::notZero, not_null);
457 stop("non-null oop required");
458 bind(not_null);
459 verify_oop(r);
460 }
461
462 void C1_MacroAssembler::invalidate_registers(bool inv_rax, bool inv_rbx, bool inv_rcx, bool inv_rdx, bool inv_rsi, bool inv_rdi) {
463 #ifdef ASSERT
464 if (inv_rax) movptr(rax, 0xDEAD);
465 if (inv_rbx) movptr(rbx, 0xDEAD);
466 if (inv_rcx) movptr(rcx, 0xDEAD);
467 if (inv_rdx) movptr(rdx, 0xDEAD);
468 if (inv_rsi) movptr(rsi, 0xDEAD);
469 if (inv_rdi) movptr(rdi, 0xDEAD);
470 #endif
471 }
472
473 #endif // ifndef PRODUCT