1 /*
2 * Copyright (c) 1999, 2024, 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 "code/compiledIC.hpp"
29 #include "compiler/compilerDefinitions.inline.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/barrierSetAssembler.hpp"
32 #include "gc/shared/collectedHeap.hpp"
33 #include "gc/shared/tlab_globals.hpp"
34 #include "interpreter/interpreter.hpp"
35 #include "oops/arrayOop.hpp"
36 #include "oops/markWord.hpp"
37 #include "runtime/basicLock.hpp"
38 #include "runtime/frame.inline.hpp"
39 #include "runtime/globals.hpp"
40 #include "runtime/os.hpp"
41 #include "runtime/sharedRuntime.hpp"
42 #include "runtime/stubRoutines.hpp"
43 #include "utilities/checkedCast.hpp"
44 #include "utilities/globalDefinitions.hpp"
45
46 int C1_MacroAssembler::lock_object(Register hdr, Register obj, Register disp_hdr, Register tmp, Label& slow_case) {
47 const int aligned_mask = BytesPerWord -1;
48 const int hdr_offset = oopDesc::mark_offset_in_bytes();
49 assert(hdr == rax, "hdr must be rax, for the cmpxchg instruction");
50 assert_different_registers(hdr, obj, disp_hdr, tmp);
51 int null_check_offset = -1;
52
53 verify_oop(obj);
54
55 // save object being locked into the BasicObjectLock
56 movptr(Address(disp_hdr, BasicObjectLock::obj_offset()), obj);
57
58 null_check_offset = offset();
59
60 if (DiagnoseSyncOnValueBasedClasses != 0) {
61 load_klass(hdr, obj, rscratch1);
62 testb(Address(hdr, Klass::misc_flags_offset()), KlassFlags::_misc_is_value_based_class);
63 jcc(Assembler::notZero, slow_case);
64 }
65
66 if (LockingMode == LM_LIGHTWEIGHT) {
67 #ifdef _LP64
68 const Register thread = r15_thread;
69 lightweight_lock(disp_hdr, obj, hdr, thread, tmp, slow_case);
70 #else
71 // Implicit null check.
72 movptr(hdr, Address(obj, oopDesc::mark_offset_in_bytes()));
73 // Lacking registers and thread on x86_32. Always take slow path.
74 jmp(slow_case);
75 #endif
76 } else if (LockingMode == LM_LEGACY) {
77 Label done;
78 // Load object header
79 movptr(hdr, Address(obj, hdr_offset));
80 // and mark it as unlocked
81 orptr(hdr, markWord::unlocked_value);
82 if (EnableValhalla) {
83 // Mask inline_type bit such that we go to the slow path if object is an inline type
84 andptr(hdr, ~((int) markWord::inline_type_bit_in_place));
85 }
86 // save unlocked object header into the displaced header location on the stack
87 movptr(Address(disp_hdr, 0), hdr);
88 // test if object header is still the same (i.e. unlocked), and if so, store the
89 // displaced header address in the object header - if it is not the same, get the
90 // object header instead
91 MacroAssembler::lock(); // must be immediately before cmpxchg!
92 cmpxchgptr(disp_hdr, Address(obj, hdr_offset));
93 // if the object header was the same, we're done
94 jcc(Assembler::equal, done);
95 // if the object header was not the same, it is now in the hdr register
96 // => test if it is a stack pointer into the same stack (recursive locking), i.e.:
97 //
98 // 1) (hdr & aligned_mask) == 0
99 // 2) rsp <= hdr
100 // 3) hdr <= rsp + page_size
101 //
102 // these 3 tests can be done by evaluating the following expression:
103 //
104 // (hdr - rsp) & (aligned_mask - page_size)
105 //
106 // assuming both the stack pointer and page_size have their least
107 // significant 2 bits cleared and page_size is a power of 2
108 subptr(hdr, rsp);
109 andptr(hdr, aligned_mask - (int)os::vm_page_size());
110 // for recursive locking, the result is zero => save it in the displaced header
111 // location (null in the displaced hdr location indicates recursive locking)
112 movptr(Address(disp_hdr, 0), hdr);
113 // otherwise we don't care about the result and handle locking via runtime call
114 jcc(Assembler::notZero, slow_case);
115 // done
116 bind(done);
117 }
118
119 inc_held_monitor_count();
120
121 return null_check_offset;
122 }
123
124 void C1_MacroAssembler::unlock_object(Register hdr, Register obj, Register disp_hdr, Label& slow_case) {
125 const int aligned_mask = BytesPerWord -1;
126 const int hdr_offset = oopDesc::mark_offset_in_bytes();
127 assert(disp_hdr == rax, "disp_hdr must be rax, for the cmpxchg instruction");
128 assert(hdr != obj && hdr != disp_hdr && obj != disp_hdr, "registers must be different");
129 Label done;
130
131 if (LockingMode != LM_LIGHTWEIGHT) {
132 // load displaced header
133 movptr(hdr, Address(disp_hdr, 0));
134 // if the loaded hdr is null we had recursive locking
135 testptr(hdr, hdr);
136 // if we had recursive locking, we are done
137 jcc(Assembler::zero, done);
138 }
139
140 // load object
141 movptr(obj, Address(disp_hdr, BasicObjectLock::obj_offset()));
142 verify_oop(obj);
143
144 if (LockingMode == LM_LIGHTWEIGHT) {
145 #ifdef _LP64
146 lightweight_unlock(obj, disp_hdr, r15_thread, hdr, slow_case);
147 #else
148 // Lacking registers and thread on x86_32. Always take slow path.
149 jmp(slow_case);
150 #endif
151 } else if (LockingMode == LM_LEGACY) {
152 // test if object header is pointing to the displaced header, and if so, restore
153 // the displaced header in the object - if the object header is not pointing to
154 // the displaced header, get the object header instead
155 MacroAssembler::lock(); // must be immediately before cmpxchg!
156 cmpxchgptr(hdr, Address(obj, hdr_offset));
157 // if the object header was not pointing to the displaced header,
158 // we do unlocking via runtime call
159 jcc(Assembler::notEqual, slow_case);
160 // done
161 }
162 bind(done);
163 dec_held_monitor_count();
164 }
165
166
167 // Defines obj, preserves var_size_in_bytes
168 void C1_MacroAssembler::try_allocate(Register obj, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2, Label& slow_case) {
169 if (UseTLAB) {
170 tlab_allocate(noreg, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
171 } else {
172 jmp(slow_case);
173 }
174 }
175
176
177 void C1_MacroAssembler::initialize_header(Register obj, Register klass, Register len, Register t1, Register t2) {
178 assert_different_registers(obj, klass, len);
179 if (EnableValhalla) {
180 // Need to copy markWord::prototype header for klass
181 assert_different_registers(obj, klass, len, t1, t2);
182 movptr(t1, Address(klass, Klass::prototype_header_offset()));
183 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), t1);
184 } else {
185 // This assumes that all prototype bits fit in an int32_t
186 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), checked_cast<int32_t>(markWord::prototype().value()));
187 }
188 #ifdef _LP64
189 if (UseCompressedClassPointers) { // Take care not to kill klass
190 movptr(t1, klass);
191 encode_klass_not_null(t1, rscratch1);
192 movl(Address(obj, oopDesc::klass_offset_in_bytes()), t1);
193 } else
194 #endif
195 {
196 movptr(Address(obj, oopDesc::klass_offset_in_bytes()), klass);
197 }
198
199 if (len->is_valid()) {
200 movl(Address(obj, arrayOopDesc::length_offset_in_bytes()), len);
201 #ifdef _LP64
202 int base_offset = arrayOopDesc::length_offset_in_bytes() + BytesPerInt;
203 if (!is_aligned(base_offset, BytesPerWord)) {
204 assert(is_aligned(base_offset, BytesPerInt), "must be 4-byte aligned");
205 // Clear gap/first 4 bytes following the length field.
206 xorl(t1, t1);
207 movl(Address(obj, base_offset), t1);
208 }
209 #endif
210 }
211 #ifdef _LP64
212 else if (UseCompressedClassPointers) {
213 xorptr(t1, t1);
214 store_klass_gap(obj, t1);
215 }
216 #endif
217 }
218
219
220 // preserves obj, destroys len_in_bytes
221 void C1_MacroAssembler::initialize_body(Register obj, Register len_in_bytes, int hdr_size_in_bytes, Register t1) {
222 assert(hdr_size_in_bytes >= 0, "header size must be positive or 0");
223 Label done;
224
225 // len_in_bytes is positive and ptr sized
226 subptr(len_in_bytes, hdr_size_in_bytes);
227 zero_memory(obj, len_in_bytes, hdr_size_in_bytes, t1);
228 bind(done);
229 }
230
231
232 void C1_MacroAssembler::allocate_object(Register obj, Register t1, Register t2, int header_size, int object_size, Register klass, Label& slow_case) {
233 assert(obj == rax, "obj must be in rax, for cmpxchg");
234 assert_different_registers(obj, t1, t2); // XXX really?
235 assert(header_size >= 0 && object_size >= header_size, "illegal sizes");
236
237 try_allocate(obj, noreg, object_size * BytesPerWord, t1, t2, slow_case);
238
239 initialize_object(obj, klass, noreg, object_size * HeapWordSize, t1, t2, UseTLAB);
240 }
241
242 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) {
243 assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0,
244 "con_size_in_bytes is not multiple of alignment");
245 const int hdr_size_in_bytes = instanceOopDesc::header_size() * HeapWordSize;
246
247 initialize_header(obj, klass, noreg, t1, t2);
248
249 if (!(UseTLAB && ZeroTLAB && is_tlab_allocated)) {
250 // clear rest of allocated space
251 const Register t1_zero = t1;
252 const Register index = t2;
253 const int threshold = 6 * BytesPerWord; // approximate break even point for code size (see comments below)
254 if (var_size_in_bytes != noreg) {
255 mov(index, var_size_in_bytes);
256 initialize_body(obj, index, hdr_size_in_bytes, t1_zero);
257 } else if (con_size_in_bytes <= threshold) {
258 // use explicit null stores
259 // code size = 2 + 3*n bytes (n = number of fields to clear)
260 xorptr(t1_zero, t1_zero); // use t1_zero reg to clear memory (shorter code)
261 for (int i = hdr_size_in_bytes; i < con_size_in_bytes; i += BytesPerWord)
262 movptr(Address(obj, i), t1_zero);
263 } else if (con_size_in_bytes > hdr_size_in_bytes) {
264 // use loop to null out the fields
265 // code size = 16 bytes for even n (n = number of fields to clear)
266 // initialize last object field first if odd number of fields
267 xorptr(t1_zero, t1_zero); // use t1_zero reg to clear memory (shorter code)
268 movptr(index, (con_size_in_bytes - hdr_size_in_bytes) >> 3);
269 // initialize last object field if constant size is odd
270 if (((con_size_in_bytes - hdr_size_in_bytes) & 4) != 0)
271 movptr(Address(obj, con_size_in_bytes - (1*BytesPerWord)), t1_zero);
272 // initialize remaining object fields: rdx is a multiple of 2
273 { Label loop;
274 bind(loop);
275 movptr(Address(obj, index, Address::times_8, hdr_size_in_bytes - (1*BytesPerWord)),
276 t1_zero);
277 NOT_LP64(movptr(Address(obj, index, Address::times_8, hdr_size_in_bytes - (2*BytesPerWord)),
278 t1_zero);)
279 decrement(index);
280 jcc(Assembler::notZero, loop);
281 }
282 }
283 }
284
285 if (CURRENT_ENV->dtrace_alloc_probes()) {
286 assert(obj == rax, "must be");
287 call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id)));
288 }
289
290 verify_oop(obj);
291 }
292
293 void C1_MacroAssembler::allocate_array(Register obj, Register len, Register t1, Register t2, int base_offset_in_bytes, Address::ScaleFactor f, Register klass, Label& slow_case, bool zero_array) {
294 assert(obj == rax, "obj must be in rax, for cmpxchg");
295 assert_different_registers(obj, len, t1, t2, klass);
296
297 // determine alignment mask
298 assert(!(BytesPerWord & 1), "must be a multiple of 2 for masking code to work");
299
300 // check for negative or excessive length
301 cmpptr(len, checked_cast<int32_t>(max_array_allocation_length));
302 jcc(Assembler::above, slow_case);
303
304 const Register arr_size = t2; // okay to be the same
305 // align object end
306 movptr(arr_size, base_offset_in_bytes + MinObjAlignmentInBytesMask);
307 lea(arr_size, Address(arr_size, len, f));
308 andptr(arr_size, ~MinObjAlignmentInBytesMask);
309
310 try_allocate(obj, arr_size, 0, t1, t2, slow_case);
311
312 initialize_header(obj, klass, len, t1, t2);
313
314 // clear rest of allocated space
315 if (zero_array) {
316 const Register len_zero = len;
317 // Align-up to word boundary, because we clear the 4 bytes potentially
318 // following the length field in initialize_header().
319 int base_offset = align_up(base_offset_in_bytes, BytesPerWord);
320 initialize_body(obj, arr_size, base_offset, len_zero);
321 }
322
323 if (CURRENT_ENV->dtrace_alloc_probes()) {
324 assert(obj == rax, "must be");
325 call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id)));
326 }
327
328 verify_oop(obj);
329 }
330
331 void C1_MacroAssembler::build_frame_helper(int frame_size_in_bytes, int sp_offset_for_orig_pc, int sp_inc, bool reset_orig_pc, bool needs_stack_repair) {
332 push(rbp);
333 if (PreserveFramePointer) {
334 mov(rbp, rsp);
335 }
336 #if !defined(_LP64) && defined(COMPILER2)
337 if (UseSSE < 2 && !CompilerConfig::is_c1_only_no_jvmci()) {
338 // c2 leaves fpu stack dirty. Clean it on entry
339 empty_FPU_stack();
340 }
341 #endif // !_LP64 && COMPILER2
342 decrement(rsp, frame_size_in_bytes);
343
344 if (needs_stack_repair) {
345 // Save stack increment (also account for fixed framesize and rbp)
346 assert((sp_inc & (StackAlignmentInBytes-1)) == 0, "stack increment not aligned");
347 int real_frame_size = sp_inc + frame_size_in_bytes + wordSize;
348 movptr(Address(rsp, frame_size_in_bytes - wordSize), real_frame_size);
349 }
350 if (reset_orig_pc) {
351 // Zero orig_pc to detect deoptimization during buffering in the entry points
352 movptr(Address(rsp, sp_offset_for_orig_pc), 0);
353 }
354 }
355
356 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) {
357 // Make sure there is enough stack space for this method's activation.
358 // Note that we do this before doing an enter(). This matches the
359 // ordering of C2's stack overflow check / rsp decrement and allows
360 // the SharedRuntime stack overflow handling to be consistent
361 // between the two compilers.
362 assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
363 generate_stack_overflow_check(bang_size_in_bytes);
364
365 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, 0, has_scalarized_args, needs_stack_repair);
366
367 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
368 // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
369 bs->nmethod_entry_barrier(this, nullptr /* slow_path */, nullptr /* continuation */);
370
371 if (verified_inline_entry_label != nullptr) {
372 // Jump here from the scalarized entry points that already created the frame.
373 bind(*verified_inline_entry_label);
374 }
375 }
376
377 void C1_MacroAssembler::verified_entry(bool breakAtEntry) {
378 if (breakAtEntry || VerifyFPU) {
379 // Verified Entry first instruction should be 5 bytes long for correct
380 // patching by patch_verified_entry().
381 //
382 // Breakpoint and VerifyFPU have one byte first instruction.
383 // Also first instruction will be one byte "push(rbp)" if stack banging
384 // code is not generated (see build_frame() above).
385 // For all these cases generate long instruction first.
386 fat_nop();
387 }
388 if (breakAtEntry) int3();
389 // build frame
390 IA32_ONLY( verify_FPU(0, "method_entry"); )
391 }
392
393 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) {
394 assert(InlineTypePassFieldsAsArgs, "sanity");
395 // Make sure there is enough stack space for this method's activation.
396 assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
397 generate_stack_overflow_check(bang_size_in_bytes);
398
399 GrowableArray<SigEntry>* sig = ces->sig();
400 GrowableArray<SigEntry>* sig_cc = is_inline_ro_entry ? ces->sig_cc_ro() : ces->sig_cc();
401 VMRegPair* regs = ces->regs();
402 VMRegPair* regs_cc = is_inline_ro_entry ? ces->regs_cc_ro() : ces->regs_cc();
403 int args_on_stack = ces->args_on_stack();
404 int args_on_stack_cc = is_inline_ro_entry ? ces->args_on_stack_cc_ro() : ces->args_on_stack_cc();
405
406 assert(sig->length() <= sig_cc->length(), "Zero-sized inline class not allowed!");
407 BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, sig_cc->length());
408 int args_passed = sig->length();
409 int args_passed_cc = SigEntry::fill_sig_bt(sig_cc, sig_bt);
410
411 // Create a temp frame so we can call into the runtime. It must be properly set up to accommodate GC.
412 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, 0, true, ces->c1_needs_stack_repair());
413
414 // The runtime call might safepoint, make sure nmethod entry barrier is executed
415 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
416 // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
417 bs->nmethod_entry_barrier(this, nullptr /* slow_path */, nullptr /* continuation */);
418
419 // FIXME -- call runtime only if we cannot in-line allocate all the incoming inline type args.
420 movptr(rbx, (intptr_t)(ces->method()));
421 if (is_inline_ro_entry) {
422 call(RuntimeAddress(Runtime1::entry_for(Runtime1::buffer_inline_args_no_receiver_id)));
423 } else {
424 call(RuntimeAddress(Runtime1::entry_for(Runtime1::buffer_inline_args_id)));
425 }
426 int rt_call_offset = offset();
427
428 // Remove the temp frame
429 addptr(rsp, frame_size_in_bytes);
430 pop(rbp);
431
432 // Check if we need to extend the stack for packing
433 int sp_inc = 0;
434 if (args_on_stack > args_on_stack_cc) {
435 sp_inc = extend_stack_for_inline_args(args_on_stack);
436 }
437
438 shuffle_inline_args(true, is_inline_ro_entry, sig_cc,
439 args_passed_cc, args_on_stack_cc, regs_cc, // from
440 args_passed, args_on_stack, regs, // to
441 sp_inc, rax);
442
443 // Create the real frame. Below jump will then skip over the stack banging and frame
444 // setup code in the verified_inline_entry (which has a different real_frame_size).
445 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, sp_inc, false, ces->c1_needs_stack_repair());
446
447 jmp(verified_inline_entry_label);
448 return rt_call_offset;
449 }
450
451 void C1_MacroAssembler::load_parameter(int offset_in_words, Register reg) {
452 // rbp, + 0: link
453 // + 1: return address
454 // + 2: argument with offset 0
455 // + 3: argument with offset 1
456 // + 4: ...
457
458 movptr(reg, Address(rbp, (offset_in_words + 2) * BytesPerWord));
459 }
460
461 #ifndef PRODUCT
462
463 void C1_MacroAssembler::verify_stack_oop(int stack_offset) {
464 if (!VerifyOops) return;
465 verify_oop_addr(Address(rsp, stack_offset));
466 }
467
468 void C1_MacroAssembler::verify_not_null_oop(Register r) {
469 if (!VerifyOops) return;
470 Label not_null;
471 testptr(r, r);
472 jcc(Assembler::notZero, not_null);
473 stop("non-null oop required");
474 bind(not_null);
475 verify_oop(r);
476 }
477
478 void C1_MacroAssembler::invalidate_registers(bool inv_rax, bool inv_rbx, bool inv_rcx, bool inv_rdx, bool inv_rsi, bool inv_rdi) {
479 #ifdef ASSERT
480 if (inv_rax) movptr(rax, 0xDEAD);
481 if (inv_rbx) movptr(rbx, 0xDEAD);
482 if (inv_rcx) movptr(rcx, 0xDEAD);
483 if (inv_rdx) movptr(rdx, 0xDEAD);
484 if (inv_rsi) movptr(rsi, 0xDEAD);
485 if (inv_rdi) movptr(rdi, 0xDEAD);
486 #endif
487 }
488
489 #endif // ifndef PRODUCT