1 /*
2 * Copyright (c) 1999, 2023, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2014, 2021, Red Hat Inc. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27 #include "c1/c1_MacroAssembler.hpp"
28 #include "c1/c1_Runtime1.hpp"
29 #include "gc/shared/barrierSetAssembler.hpp"
30 #include "gc/shared/collectedHeap.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/barrierSetAssembler.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/os.hpp"
39 #include "runtime/sharedRuntime.hpp"
40 #include "runtime/stubRoutines.hpp"
41
42 void C1_MacroAssembler::float_cmp(bool is_float, int unordered_result,
43 FloatRegister f0, FloatRegister f1,
44 Register result)
45 {
46 Label done;
47 if (is_float) {
48 fcmps(f0, f1);
49 } else {
50 fcmpd(f0, f1);
51 }
52 if (unordered_result < 0) {
53 // we want -1 for unordered or less than, 0 for equal and 1 for
54 // greater than.
55 cset(result, NE); // Not equal or unordered
56 cneg(result, result, LT); // Less than or unordered
57 } else {
58 // we want -1 for less than, 0 for equal and 1 for unordered or
59 // greater than.
60 cset(result, NE); // Not equal or unordered
61 cneg(result, result, LO); // Less than
62 }
63 }
64
65 int C1_MacroAssembler::lock_object(Register hdr, Register obj, Register disp_hdr, Register temp, Label& slow_case) {
66 const int aligned_mask = BytesPerWord -1;
67 const int hdr_offset = oopDesc::mark_offset_in_bytes();
68 assert_different_registers(hdr, obj, disp_hdr, temp, rscratch2);
69 int null_check_offset = -1;
70
71 verify_oop(obj);
72
73 // save object being locked into the BasicObjectLock
74 str(obj, Address(disp_hdr, BasicObjectLock::obj_offset()));
75
76 null_check_offset = offset();
77
78 if (DiagnoseSyncOnValueBasedClasses != 0) {
79 load_klass(hdr, obj);
80 ldrw(hdr, Address(hdr, Klass::access_flags_offset()));
81 tstw(hdr, JVM_ACC_IS_VALUE_BASED_CLASS);
82 br(Assembler::NE, slow_case);
83 }
84
85 // Load object header
86 ldr(hdr, Address(obj, hdr_offset));
87 if (LockingMode == LM_LIGHTWEIGHT) {
88 lightweight_lock(obj, hdr, temp, rscratch2, slow_case);
89 } else if (LockingMode == LM_LEGACY) {
90 Label done;
91 // and mark it as unlocked
92 orr(hdr, hdr, markWord::unlocked_value);
93
94 if (EnableValhalla) {
95 // Mask always_locked bit such that we go to the slow path if object is an inline type
96 andr(hdr, hdr, ~markWord::inline_type_bit_in_place);
97 }
98
99 // save unlocked object header into the displaced header location on the stack
100 str(hdr, Address(disp_hdr, 0));
101 // test if object header is still the same (i.e. unlocked), and if so, store the
102 // displaced header address in the object header - if it is not the same, get the
103 // object header instead
104 lea(rscratch2, Address(obj, hdr_offset));
105 cmpxchgptr(hdr, disp_hdr, rscratch2, rscratch1, done, /*fallthough*/nullptr);
106 // if the object header was the same, we're done
107 // if the object header was not the same, it is now in the hdr register
108 // => test if it is a stack pointer into the same stack (recursive locking), i.e.:
109 //
110 // 1) (hdr & aligned_mask) == 0
111 // 2) sp <= hdr
112 // 3) hdr <= sp + page_size
113 //
114 // these 3 tests can be done by evaluating the following expression:
115 //
116 // (hdr - sp) & (aligned_mask - page_size)
117 //
118 // assuming both the stack pointer and page_size have their least
119 // significant 2 bits cleared and page_size is a power of 2
120 mov(rscratch1, sp);
121 sub(hdr, hdr, rscratch1);
122 ands(hdr, hdr, aligned_mask - (int)os::vm_page_size());
123 // for recursive locking, the result is zero => save it in the displaced header
124 // location (null in the displaced hdr location indicates recursive locking)
125 str(hdr, Address(disp_hdr, 0));
126 // otherwise we don't care about the result and handle locking via runtime call
127 cbnz(hdr, slow_case);
128 // done
129 bind(done);
130 }
131 increment(Address(rthread, JavaThread::held_monitor_count_offset()));
132 return null_check_offset;
133 }
134
135
136 void C1_MacroAssembler::unlock_object(Register hdr, Register obj, Register disp_hdr, Register temp, Label& slow_case) {
137 const int aligned_mask = BytesPerWord -1;
138 const int hdr_offset = oopDesc::mark_offset_in_bytes();
139 assert_different_registers(hdr, obj, disp_hdr, temp, rscratch2);
140 Label done;
141
142 if (LockingMode != LM_LIGHTWEIGHT) {
143 // load displaced header
144 ldr(hdr, Address(disp_hdr, 0));
145 // if the loaded hdr is null we had recursive locking
146 // if we had recursive locking, we are done
147 cbz(hdr, done);
148 }
149
150 // load object
151 ldr(obj, Address(disp_hdr, BasicObjectLock::obj_offset()));
152 verify_oop(obj);
153
154 if (LockingMode == LM_LIGHTWEIGHT) {
155 ldr(hdr, Address(obj, oopDesc::mark_offset_in_bytes()));
156 // We cannot use tbnz here, the target might be too far away and cannot
157 // be encoded.
158 tst(hdr, markWord::monitor_value);
159 br(Assembler::NE, slow_case);
160 lightweight_unlock(obj, hdr, temp, rscratch2, slow_case);
161 } else if (LockingMode == LM_LEGACY) {
162 // test if object header is pointing to the displaced header, and if so, restore
163 // the displaced header in the object - if the object header is not pointing to
164 // the displaced header, get the object header instead
165 // if the object header was not pointing to the displaced header,
166 // we do unlocking via runtime call
167 if (hdr_offset) {
168 lea(rscratch1, Address(obj, hdr_offset));
169 cmpxchgptr(disp_hdr, hdr, rscratch1, rscratch2, done, &slow_case);
170 } else {
171 cmpxchgptr(disp_hdr, hdr, obj, rscratch2, done, &slow_case);
172 }
173 // done
174 bind(done);
175 }
176 decrement(Address(rthread, JavaThread::held_monitor_count_offset()));
177 }
178
179
180 // Defines obj, preserves var_size_in_bytes
181 void C1_MacroAssembler::try_allocate(Register obj, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2, Label& slow_case) {
182 if (UseTLAB) {
183 tlab_allocate(obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
184 } else {
185 b(slow_case);
186 }
187 }
188
189 void C1_MacroAssembler::initialize_header(Register obj, Register klass, Register len, Register t1, Register t2) {
190 assert_different_registers(obj, klass, len);
191 if (EnableValhalla) {
192 // Need to copy markWord::prototype header for klass
193 assert_different_registers(obj, klass, len, t1, t2);
194 ldr(t1, Address(klass, Klass::prototype_header_offset()));
195 } else {
196 // This assumes that all prototype bits fit in an int32_t
197 mov(t1, (int32_t)(intptr_t)markWord::prototype().value());
198 }
199 str(t1, Address(obj, oopDesc::mark_offset_in_bytes()));
200
201 if (UseCompressedClassPointers) { // Take care not to kill klass
202 encode_klass_not_null(t1, klass);
203 strw(t1, Address(obj, oopDesc::klass_offset_in_bytes()));
204 } else {
205 str(klass, Address(obj, oopDesc::klass_offset_in_bytes()));
206 }
207
208 if (len->is_valid()) {
209 strw(len, Address(obj, arrayOopDesc::length_offset_in_bytes()));
210 } else if (UseCompressedClassPointers) {
211 store_klass_gap(obj, zr);
212 }
213 }
214
215 // preserves obj, destroys len_in_bytes
216 //
217 // Scratch registers: t1 = r10, t2 = r11
218 //
219 void C1_MacroAssembler::initialize_body(Register obj, Register len_in_bytes, int hdr_size_in_bytes, Register t1, Register t2) {
220 assert(hdr_size_in_bytes >= 0, "header size must be positive or 0");
221 assert(t1 == r10 && t2 == r11, "must be");
222
223 Label done;
224
225 // len_in_bytes is positive and ptr sized
226 subs(len_in_bytes, len_in_bytes, hdr_size_in_bytes);
227 br(Assembler::EQ, done);
228
229 // zero_words() takes ptr in r10 and count in words in r11
230 mov(rscratch1, len_in_bytes);
231 lea(t1, Address(obj, hdr_size_in_bytes));
232 lsr(t2, rscratch1, LogBytesPerWord);
233 address tpc = zero_words(t1, t2);
234
235 bind(done);
236 if (tpc == nullptr) {
237 Compilation::current()->bailout("no space for trampoline stub");
238 }
239 }
240
241
242 void C1_MacroAssembler::allocate_object(Register obj, Register t1, Register t2, int header_size, int object_size, Register klass, Label& slow_case) {
243 assert_different_registers(obj, t1, t2); // XXX really?
244 assert(header_size >= 0 && object_size >= header_size, "illegal sizes");
245
246 try_allocate(obj, noreg, object_size * BytesPerWord, t1, t2, slow_case);
247
248 initialize_object(obj, klass, noreg, object_size * HeapWordSize, t1, t2, UseTLAB);
249 }
250
251 // Scratch registers: t1 = r10, t2 = r11
252 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) {
253 assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0,
254 "con_size_in_bytes is not multiple of alignment");
255 const int hdr_size_in_bytes = instanceOopDesc::header_size() * HeapWordSize;
256
257 initialize_header(obj, klass, noreg, t1, t2);
258
259 if (!(UseTLAB && ZeroTLAB && is_tlab_allocated)) {
260 // clear rest of allocated space
261 const Register index = t2;
262 if (var_size_in_bytes != noreg) {
263 mov(index, var_size_in_bytes);
264 initialize_body(obj, index, hdr_size_in_bytes, t1, t2);
265 if (Compilation::current()->bailed_out()) {
266 return;
267 }
268 } else if (con_size_in_bytes > hdr_size_in_bytes) {
269 con_size_in_bytes -= hdr_size_in_bytes;
270 lea(t1, Address(obj, hdr_size_in_bytes));
271 address tpc = zero_words(t1, con_size_in_bytes / BytesPerWord);
272 if (tpc == nullptr) {
273 Compilation::current()->bailout("no space for trampoline stub");
274 return;
275 }
276 }
277 }
278
279 membar(StoreStore);
280
281 if (CURRENT_ENV->dtrace_alloc_probes()) {
282 assert(obj == r0, "must be");
283 far_call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id)));
284 }
285
286 verify_oop(obj);
287 }
288
289 void C1_MacroAssembler::allocate_array(Register obj, Register len, Register t1, Register t2, int header_size, int f, Register klass, Label& slow_case) {
290 assert_different_registers(obj, len, t1, t2, klass);
291
292 // determine alignment mask
293 assert(!(BytesPerWord & 1), "must be a multiple of 2 for masking code to work");
294
295 // check for negative or excessive length
296 mov(rscratch1, (int32_t)max_array_allocation_length);
297 cmp(len, rscratch1);
298 br(Assembler::HS, slow_case);
299
300 const Register arr_size = t2; // okay to be the same
301 // align object end
302 mov(arr_size, (int32_t)header_size * BytesPerWord + MinObjAlignmentInBytesMask);
303 add(arr_size, arr_size, len, ext::uxtw, f);
304 andr(arr_size, arr_size, ~MinObjAlignmentInBytesMask);
305
306 try_allocate(obj, arr_size, 0, t1, t2, slow_case);
307
308 initialize_header(obj, klass, len, t1, t2);
309
310 // clear rest of allocated space
311 initialize_body(obj, arr_size, header_size * BytesPerWord, t1, t2);
312 if (Compilation::current()->bailed_out()) {
313 return;
314 }
315
316 membar(StoreStore);
317
318 if (CURRENT_ENV->dtrace_alloc_probes()) {
319 assert(obj == r0, "must be");
320 far_call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id)));
321 }
322
323 verify_oop(obj);
324 }
325
326
327 void C1_MacroAssembler::inline_cache_check(Register receiver, Register iCache) {
328 verify_oop(receiver);
329 // explicit null check not needed since load from [klass_offset] causes a trap
330 // check against inline cache
331 assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()), "must add explicit null check");
332
333 cmp_klass(receiver, iCache, rscratch1);
334 }
335
336 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) {
337 MacroAssembler::build_frame(frame_size_in_bytes);
338
339 if (needs_stack_repair) {
340 save_stack_increment(sp_inc, frame_size_in_bytes);
341 }
342 if (reset_orig_pc) {
343 // Zero orig_pc to detect deoptimization during buffering in the entry points
344 str(zr, Address(sp, sp_offset_for_orig_pc));
345 }
346 }
347
348 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) {
349 // Make sure there is enough stack space for this method's activation.
350 // Note that we do this before creating a frame.
351 assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
352 generate_stack_overflow_check(bang_size_in_bytes);
353
354 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, 0, has_scalarized_args, needs_stack_repair);
355
356 // Insert nmethod entry barrier into frame.
357 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
358 bs->nmethod_entry_barrier(this, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
359
360 if (verified_inline_entry_label != nullptr) {
361 // Jump here from the scalarized entry points that already created the frame.
362 bind(*verified_inline_entry_label);
363 }
364 }
365
366 void C1_MacroAssembler::verified_entry(bool breakAtEntry) {
367 // If we have to make this method not-entrant we'll overwrite its
368 // first instruction with a jump. For this action to be legal we
369 // must ensure that this first instruction is a B, BL, NOP, BKPT,
370 // SVC, HVC, or SMC. Make it a NOP.
371 nop();
372 if (C1Breakpoint) brk(1);
373 }
374
375 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) {
376 assert(InlineTypePassFieldsAsArgs, "sanity");
377 // Make sure there is enough stack space for this method's activation.
378 assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
379 generate_stack_overflow_check(bang_size_in_bytes);
380
381 GrowableArray<SigEntry>* sig = ces->sig();
382 GrowableArray<SigEntry>* sig_cc = is_inline_ro_entry ? ces->sig_cc_ro() : ces->sig_cc();
383 VMRegPair* regs = ces->regs();
384 VMRegPair* regs_cc = is_inline_ro_entry ? ces->regs_cc_ro() : ces->regs_cc();
385 int args_on_stack = ces->args_on_stack();
386 int args_on_stack_cc = is_inline_ro_entry ? ces->args_on_stack_cc_ro() : ces->args_on_stack_cc();
387
388 assert(sig->length() <= sig_cc->length(), "Zero-sized inline class not allowed!");
389 BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, sig_cc->length());
390 int args_passed = sig->length();
391 int args_passed_cc = SigEntry::fill_sig_bt(sig_cc, sig_bt);
392
393 // Create a temp frame so we can call into the runtime. It must be properly set up to accommodate GC.
394 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, 0, true, ces->c1_needs_stack_repair());
395
396 // The runtime call might safepoint, make sure nmethod entry barrier is executed
397 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
398 // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
399 bs->nmethod_entry_barrier(this, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
400
401 // FIXME -- call runtime only if we cannot in-line allocate all the incoming inline type args.
402 mov(r19, (intptr_t) ces->method());
403 if (is_inline_ro_entry) {
404 far_call(RuntimeAddress(Runtime1::entry_for(Runtime1::buffer_inline_args_no_receiver_id)));
405 } else {
406 far_call(RuntimeAddress(Runtime1::entry_for(Runtime1::buffer_inline_args_id)));
407 }
408 int rt_call_offset = offset();
409
410 // The runtime call returns the new array in r20 instead of the usual r0
411 // because r0 is also j_rarg7 which may be holding a live argument here.
412 Register val_array = r20;
413
414 // Remove the temp frame
415 MacroAssembler::remove_frame(frame_size_in_bytes);
416
417 // Check if we need to extend the stack for packing
418 int sp_inc = 0;
419 if (args_on_stack > args_on_stack_cc) {
420 sp_inc = extend_stack_for_inline_args(args_on_stack);
421 }
422
423 shuffle_inline_args(true, is_inline_ro_entry, sig_cc,
424 args_passed_cc, args_on_stack_cc, regs_cc, // from
425 args_passed, args_on_stack, regs, // to
426 sp_inc, val_array);
427
428 // Create the real frame. Below jump will then skip over the stack banging and frame
429 // setup code in the verified_inline_entry (which has a different real_frame_size).
430 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, sp_inc, false, ces->c1_needs_stack_repair());
431
432 b(verified_inline_entry_label);
433 return rt_call_offset;
434 }
435
436
437 void C1_MacroAssembler::load_parameter(int offset_in_words, Register reg) {
438 // rfp, + 0: link
439 // + 1: return address
440 // + 2: argument with offset 0
441 // + 3: argument with offset 1
442 // + 4: ...
443
444 ldr(reg, Address(rfp, (offset_in_words + 2) * BytesPerWord));
445 }
446
447 #ifndef PRODUCT
448
449 void C1_MacroAssembler::verify_stack_oop(int stack_offset) {
450 if (!VerifyOops) return;
451 verify_oop_addr(Address(sp, stack_offset));
452 }
453
454 void C1_MacroAssembler::verify_not_null_oop(Register r) {
455 if (!VerifyOops) return;
456 Label not_null;
457 cbnz(r, not_null);
458 stop("non-null oop required");
459 bind(not_null);
460 verify_oop(r);
461 }
462
463 void C1_MacroAssembler::invalidate_registers(bool inv_r0, bool inv_r19, bool inv_r2, bool inv_r3, bool inv_r4, bool inv_r5) {
464 #ifdef ASSERT
465 static int nn;
466 if (inv_r0) mov(r0, 0xDEAD);
467 if (inv_r19) mov(r19, 0xDEAD);
468 if (inv_r2) mov(r2, nn++);
469 if (inv_r3) mov(r3, 0xDEAD);
470 if (inv_r4) mov(r4, 0xDEAD);
471 if (inv_r5) mov(r5, 0xDEAD);
472 #endif
473 }
474 #endif // ifndef PRODUCT