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/os.hpp"
38 #include "runtime/sharedRuntime.hpp"
39 #include "runtime/stubRoutines.hpp"
40
41 int C1_MacroAssembler::lock_object(Register hdr, Register obj, Register disp_hdr, Register tmp, Label& slow_case) {
42 const int aligned_mask = BytesPerWord -1;
43 const int hdr_offset = oopDesc::mark_offset_in_bytes();
44 assert(hdr == rax, "hdr must be rax, for the cmpxchg instruction");
45 assert_different_registers(hdr, obj, disp_hdr, tmp);
46 int null_check_offset = -1;
47
48 verify_oop(obj);
49
50 // save object being locked into the BasicObjectLock
51 movptr(Address(disp_hdr, BasicObjectLock::obj_offset()), obj);
52
53 null_check_offset = offset();
54
55 if (DiagnoseSyncOnValueBasedClasses != 0) {
56 load_klass(hdr, obj, rscratch1);
57 movl(hdr, Address(hdr, Klass::access_flags_offset()));
58 testl(hdr, JVM_ACC_IS_VALUE_BASED_CLASS);
59 jcc(Assembler::notZero, slow_case);
60 }
61
62 // Load object header
63 movptr(hdr, Address(obj, hdr_offset));
64
65 if (LockingMode == LM_LIGHTWEIGHT) {
66 #ifdef _LP64
67 const Register thread = r15_thread;
68 #else
69 const Register thread = disp_hdr;
70 get_thread(thread);
71 #endif
72 fast_lock_impl(obj, hdr, thread, tmp, slow_case);
73 } else if (LockingMode == LM_LEGACY) {
74 Label done;
75 // and mark it as unlocked
76 orptr(hdr, markWord::unlocked_value);
77 // save unlocked object header into the displaced header location on the stack
78 movptr(Address(disp_hdr, 0), hdr);
79 // test if object header is still the same (i.e. unlocked), and if so, store the
80 // displaced header address in the object header - if it is not the same, get the
81 // object header instead
82 MacroAssembler::lock(); // must be immediately before cmpxchg!
83 cmpxchgptr(disp_hdr, Address(obj, hdr_offset));
84 // if the object header was the same, we're done
85 jcc(Assembler::equal, done);
86 // if the object header was not the same, it is now in the hdr register
87 // => test if it is a stack pointer into the same stack (recursive locking), i.e.:
88 //
89 // 1) (hdr & aligned_mask) == 0
90 // 2) rsp <= hdr
91 // 3) hdr <= rsp + page_size
92 //
93 // these 3 tests can be done by evaluating the following expression:
94 //
95 // (hdr - rsp) & (aligned_mask - page_size)
96 //
147 jcc(Assembler::notEqual, slow_case);
148 // done
149 }
150 bind(done);
151 dec_held_monitor_count();
152 }
153
154
155 // Defines obj, preserves var_size_in_bytes
156 void C1_MacroAssembler::try_allocate(Register obj, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2, Label& slow_case) {
157 if (UseTLAB) {
158 tlab_allocate(noreg, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
159 } else {
160 jmp(slow_case);
161 }
162 }
163
164
165 void C1_MacroAssembler::initialize_header(Register obj, Register klass, Register len, Register t1, Register t2) {
166 assert_different_registers(obj, klass, len);
167 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), checked_cast<int32_t>(markWord::prototype().value()));
168 #ifdef _LP64
169 if (UseCompressedClassPointers) { // Take care not to kill klass
170 movptr(t1, klass);
171 encode_klass_not_null(t1, rscratch1);
172 movl(Address(obj, oopDesc::klass_offset_in_bytes()), t1);
173 } else
174 #endif
175 {
176 movptr(Address(obj, oopDesc::klass_offset_in_bytes()), klass);
177 }
178
179 if (len->is_valid()) {
180 movl(Address(obj, arrayOopDesc::length_offset_in_bytes()), len);
181 }
182 #ifdef _LP64
183 else if (UseCompressedClassPointers) {
184 xorptr(t1, t1);
185 store_klass_gap(obj, t1);
186 }
187 #endif
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, rscratch2);
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) {
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 push(rbp);
331 if (PreserveFramePointer) {
332 mov(rbp, rsp);
333 }
334 #if !defined(_LP64) && defined(COMPILER2)
335 if (UseSSE < 2 && !CompilerConfig::is_c1_only_no_jvmci()) {
336 // c2 leaves fpu stack dirty. Clean it on entry
337 empty_FPU_stack();
338 }
339 #endif // !_LP64 && COMPILER2
340 decrement(rsp, frame_size_in_bytes); // does not emit code for frame_size == 0
341
342 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
343 // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
344 bs->nmethod_entry_barrier(this, nullptr /* slow_path */, nullptr /* continuation */);
345 }
346
347
348 void C1_MacroAssembler::remove_frame(int frame_size_in_bytes) {
349 increment(rsp, frame_size_in_bytes); // Does not emit code for frame_size == 0
350 pop(rbp);
351 }
352
353
354 void C1_MacroAssembler::verified_entry(bool breakAtEntry) {
355 if (breakAtEntry || VerifyFPU) {
356 // Verified Entry first instruction should be 5 bytes long for correct
357 // patching by patch_verified_entry().
358 //
359 // Breakpoint and VerifyFPU have one byte first instruction.
360 // Also first instruction will be one byte "push(rbp)" if stack banging
361 // code is not generated (see build_frame() above).
362 // For all these cases generate long instruction first.
363 fat_nop();
364 }
365 if (breakAtEntry) int3();
366 // build frame
367 IA32_ONLY( verify_FPU(0, "method_entry"); )
368 }
369
370 void C1_MacroAssembler::load_parameter(int offset_in_words, Register reg) {
371 // rbp, + 0: link
372 // + 1: return address
373 // + 2: argument with offset 0
374 // + 3: argument with offset 1
375 // + 4: ...
376
377 movptr(reg, Address(rbp, (offset_in_words + 2) * BytesPerWord));
378 }
379
380 #ifndef PRODUCT
381
382 void C1_MacroAssembler::verify_stack_oop(int stack_offset) {
383 if (!VerifyOops) return;
384 verify_oop_addr(Address(rsp, stack_offset));
385 }
386
387 void C1_MacroAssembler::verify_not_null_oop(Register r) {
388 if (!VerifyOops) return;
389 Label not_null;
|
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, Register tmp, 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_different_registers(hdr, obj, disp_hdr, tmp);
47 int null_check_offset = -1;
48
49 verify_oop(obj);
50
51 // save object being locked into the BasicObjectLock
52 movptr(Address(disp_hdr, BasicObjectLock::obj_offset()), obj);
53
54 null_check_offset = offset();
55
56 if (DiagnoseSyncOnValueBasedClasses != 0) {
57 load_klass(hdr, obj, rscratch1);
58 movl(hdr, Address(hdr, Klass::access_flags_offset()));
59 testl(hdr, JVM_ACC_IS_VALUE_BASED_CLASS);
60 jcc(Assembler::notZero, slow_case);
61 }
62
63 // Load object header
64 movptr(hdr, Address(obj, hdr_offset));
65
66 if (LockingMode == LM_LIGHTWEIGHT) {
67 #ifdef _LP64
68 const Register thread = r15_thread;
69 #else
70 const Register thread = disp_hdr;
71 get_thread(thread);
72 #endif
73 fast_lock_impl(obj, hdr, thread, tmp, slow_case);
74 } else if (LockingMode == LM_LEGACY) {
75 Label done;
76 // and mark it as unlocked
77 orptr(hdr, markWord::unlocked_value);
78 if (EnableValhalla) {
79 // Mask inline_type bit such that we go to the slow path if object is an inline type
80 andptr(hdr, ~((int) markWord::inline_type_bit_in_place));
81 }
82 // save unlocked object header into the displaced header location on the stack
83 movptr(Address(disp_hdr, 0), hdr);
84 // test if object header is still the same (i.e. unlocked), and if so, store the
85 // displaced header address in the object header - if it is not the same, get the
86 // object header instead
87 MacroAssembler::lock(); // must be immediately before cmpxchg!
88 cmpxchgptr(disp_hdr, Address(obj, hdr_offset));
89 // if the object header was the same, we're done
90 jcc(Assembler::equal, done);
91 // if the object header was not the same, it is now in the hdr register
92 // => test if it is a stack pointer into the same stack (recursive locking), i.e.:
93 //
94 // 1) (hdr & aligned_mask) == 0
95 // 2) rsp <= hdr
96 // 3) hdr <= rsp + page_size
97 //
98 // these 3 tests can be done by evaluating the following expression:
99 //
100 // (hdr - rsp) & (aligned_mask - page_size)
101 //
152 jcc(Assembler::notEqual, slow_case);
153 // done
154 }
155 bind(done);
156 dec_held_monitor_count();
157 }
158
159
160 // Defines obj, preserves var_size_in_bytes
161 void C1_MacroAssembler::try_allocate(Register obj, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2, Label& slow_case) {
162 if (UseTLAB) {
163 tlab_allocate(noreg, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
164 } else {
165 jmp(slow_case);
166 }
167 }
168
169
170 void C1_MacroAssembler::initialize_header(Register obj, Register klass, Register len, Register t1, Register t2) {
171 assert_different_registers(obj, klass, len);
172 if (EnableValhalla) {
173 // Need to copy markWord::prototype header for klass
174 assert_different_registers(obj, klass, len, t1, t2);
175 movptr(t1, Address(klass, Klass::prototype_header_offset()));
176 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), t1);
177 } else {
178 // This assumes that all prototype bits fit in an int32_t
179 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), checked_cast<int32_t>(markWord::prototype().value()));
180 }
181 #ifdef _LP64
182 if (UseCompressedClassPointers) { // Take care not to kill klass
183 movptr(t1, klass);
184 encode_klass_not_null(t1, rscratch1);
185 movl(Address(obj, oopDesc::klass_offset_in_bytes()), t1);
186 } else
187 #endif
188 {
189 movptr(Address(obj, oopDesc::klass_offset_in_bytes()), klass);
190 }
191
192 if (len->is_valid()) {
193 movl(Address(obj, arrayOopDesc::length_offset_in_bytes()), len);
194 }
195 #ifdef _LP64
196 else if (UseCompressedClassPointers) {
197 xorptr(t1, t1);
198 store_klass_gap(obj, t1);
199 }
200 #endif
313 verify_oop(receiver);
314 // explicit null check not needed since load from [klass_offset] causes a trap
315 // check against inline cache
316 assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()), "must add explicit null check");
317 int start_offset = offset();
318
319 if (UseCompressedClassPointers) {
320 load_klass(rscratch1, receiver, rscratch2);
321 cmpptr(rscratch1, iCache);
322 } else {
323 cmpptr(iCache, Address(receiver, oopDesc::klass_offset_in_bytes()));
324 }
325 // if icache check fails, then jump to runtime routine
326 // Note: RECEIVER must still contain the receiver!
327 jump_cc(Assembler::notEqual,
328 RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
329 const int ic_cmp_size = LP64_ONLY(10) NOT_LP64(9);
330 assert(UseCompressedClassPointers || offset() - start_offset == ic_cmp_size, "check alignment in emit_method_entry");
331 }
332
333 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) {
334 push(rbp);
335 if (PreserveFramePointer) {
336 mov(rbp, rsp);
337 }
338 #if !defined(_LP64) && defined(COMPILER2)
339 if (UseSSE < 2 && !CompilerConfig::is_c1_only_no_jvmci()) {
340 // c2 leaves fpu stack dirty. Clean it on entry
341 empty_FPU_stack();
342 }
343 #endif // !_LP64 && COMPILER2
344 decrement(rsp, frame_size_in_bytes);
345
346 if (needs_stack_repair) {
347 // Save stack increment (also account for fixed framesize and rbp)
348 assert((sp_inc & (StackAlignmentInBytes-1)) == 0, "stack increment not aligned");
349 int real_frame_size = sp_inc + frame_size_in_bytes + wordSize;
350 movptr(Address(rsp, frame_size_in_bytes - wordSize), real_frame_size);
351 }
352 if (reset_orig_pc) {
353 // Zero orig_pc to detect deoptimization during buffering in the entry points
354 movptr(Address(rsp, sp_offset_for_orig_pc), 0);
355 }
356 }
357
358 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) {
359 // Make sure there is enough stack space for this method's activation.
360 // Note that we do this before doing an enter(). This matches the
361 // ordering of C2's stack overflow check / rsp decrement and allows
362 // the SharedRuntime stack overflow handling to be consistent
363 // between the two compilers.
364 assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
365 generate_stack_overflow_check(bang_size_in_bytes);
366
367 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, 0, has_scalarized_args, needs_stack_repair);
368
369 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
370 // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
371 bs->nmethod_entry_barrier(this, nullptr /* slow_path */, nullptr /* continuation */);
372
373 if (verified_inline_entry_label != nullptr) {
374 // Jump here from the scalarized entry points that already created the frame.
375 bind(*verified_inline_entry_label);
376 }
377 }
378
379 void C1_MacroAssembler::verified_entry(bool breakAtEntry) {
380 if (breakAtEntry || VerifyFPU) {
381 // Verified Entry first instruction should be 5 bytes long for correct
382 // patching by patch_verified_entry().
383 //
384 // Breakpoint and VerifyFPU have one byte first instruction.
385 // Also first instruction will be one byte "push(rbp)" if stack banging
386 // code is not generated (see build_frame() above).
387 // For all these cases generate long instruction first.
388 fat_nop();
389 }
390 if (breakAtEntry) int3();
391 // build frame
392 IA32_ONLY( verify_FPU(0, "method_entry"); )
393 }
394
395 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) {
396 assert(InlineTypePassFieldsAsArgs, "sanity");
397 // Make sure there is enough stack space for this method's activation.
398 assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
399 generate_stack_overflow_check(bang_size_in_bytes);
400
401 GrowableArray<SigEntry>* sig = ces->sig();
402 GrowableArray<SigEntry>* sig_cc = is_inline_ro_entry ? ces->sig_cc_ro() : ces->sig_cc();
403 VMRegPair* regs = ces->regs();
404 VMRegPair* regs_cc = is_inline_ro_entry ? ces->regs_cc_ro() : ces->regs_cc();
405 int args_on_stack = ces->args_on_stack();
406 int args_on_stack_cc = is_inline_ro_entry ? ces->args_on_stack_cc_ro() : ces->args_on_stack_cc();
407
408 assert(sig->length() <= sig_cc->length(), "Zero-sized inline class not allowed!");
409 BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, sig_cc->length());
410 int args_passed = sig->length();
411 int args_passed_cc = SigEntry::fill_sig_bt(sig_cc, sig_bt);
412
413 // Create a temp frame so we can call into the runtime. It must be properly set up to accommodate GC.
414 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, 0, true, ces->c1_needs_stack_repair());
415
416 // The runtime call might safepoint, make sure nmethod entry barrier is executed
417 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
418 // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
419 bs->nmethod_entry_barrier(this, nullptr /* slow_path */, nullptr /* continuation */);
420
421 // FIXME -- call runtime only if we cannot in-line allocate all the incoming inline type args.
422 movptr(rbx, (intptr_t)(ces->method()));
423 if (is_inline_ro_entry) {
424 call(RuntimeAddress(Runtime1::entry_for(Runtime1::buffer_inline_args_no_receiver_id)));
425 } else {
426 call(RuntimeAddress(Runtime1::entry_for(Runtime1::buffer_inline_args_id)));
427 }
428 int rt_call_offset = offset();
429
430 // Remove the temp frame
431 addptr(rsp, frame_size_in_bytes);
432 pop(rbp);
433
434 // Check if we need to extend the stack for packing
435 int sp_inc = 0;
436 if (args_on_stack > args_on_stack_cc) {
437 sp_inc = extend_stack_for_inline_args(args_on_stack);
438 }
439
440 shuffle_inline_args(true, is_inline_ro_entry, sig_cc,
441 args_passed_cc, args_on_stack_cc, regs_cc, // from
442 args_passed, args_on_stack, regs, // to
443 sp_inc, rax);
444
445 // Create the real frame. Below jump will then skip over the stack banging and frame
446 // setup code in the verified_inline_entry (which has a different real_frame_size).
447 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, sp_inc, false, ces->c1_needs_stack_repair());
448
449 jmp(verified_inline_entry_label);
450 return rt_call_offset;
451 }
452
453 void C1_MacroAssembler::load_parameter(int offset_in_words, Register reg) {
454 // rbp, + 0: link
455 // + 1: return address
456 // + 2: argument with offset 0
457 // + 3: argument with offset 1
458 // + 4: ...
459
460 movptr(reg, Address(rbp, (offset_in_words + 2) * BytesPerWord));
461 }
462
463 #ifndef PRODUCT
464
465 void C1_MacroAssembler::verify_stack_oop(int stack_offset) {
466 if (!VerifyOops) return;
467 verify_oop_addr(Address(rsp, stack_offset));
468 }
469
470 void C1_MacroAssembler::verify_not_null_oop(Register r) {
471 if (!VerifyOops) return;
472 Label not_null;
|