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 "c1/c1_MacroAssembler.hpp"
26 #include "c1/c1_Runtime1.hpp"
27 #include "code/compiledIC.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/globals.hpp"
38 #include "runtime/os.hpp"
39 #include "runtime/sharedRuntime.hpp"
40 #include "runtime/stubRoutines.hpp"
41 #include "utilities/checkedCast.hpp"
42 #include "utilities/globalDefinitions.hpp"
43
44 int C1_MacroAssembler::lock_object(Register hdr, Register obj, Register disp_hdr, Register tmp, Label& slow_case) {
45 const int aligned_mask = BytesPerWord -1;
46 const int hdr_offset = oopDesc::mark_offset_in_bytes();
47 assert(hdr == rax, "hdr must be rax, for the cmpxchg instruction");
48 assert_different_registers(hdr, obj, disp_hdr, tmp);
49 int null_check_offset = -1;
50
51 verify_oop(obj);
52
53 // save object being locked into the BasicObjectLock
54 movptr(Address(disp_hdr, BasicObjectLock::obj_offset()), obj);
55
56 null_check_offset = offset();
60 testb(Address(hdr, Klass::misc_flags_offset()), KlassFlags::_misc_is_value_based_class);
61 jcc(Assembler::notZero, slow_case);
62 }
63
64 if (LockingMode == LM_LIGHTWEIGHT) {
65 #ifdef _LP64
66 const Register thread = r15_thread;
67 lightweight_lock(disp_hdr, obj, hdr, thread, tmp, slow_case);
68 #else
69 // Implicit null check.
70 movptr(hdr, Address(obj, oopDesc::mark_offset_in_bytes()));
71 // Lacking registers and thread on x86_32. Always take slow path.
72 jmp(slow_case);
73 #endif
74 } else if (LockingMode == LM_LEGACY) {
75 Label done;
76 // Load object header
77 movptr(hdr, Address(obj, hdr_offset));
78 // and mark it as unlocked
79 orptr(hdr, markWord::unlocked_value);
80 // save unlocked object header into the displaced header location on the stack
81 movptr(Address(disp_hdr, 0), hdr);
82 // test if object header is still the same (i.e. unlocked), and if so, store the
83 // displaced header address in the object header - if it is not the same, get the
84 // object header instead
85 MacroAssembler::lock(); // must be immediately before cmpxchg!
86 cmpxchgptr(disp_hdr, Address(obj, hdr_offset));
87 // if the object header was the same, we're done
88 jcc(Assembler::equal, done);
89 // if the object header was not the same, it is now in the hdr register
90 // => test if it is a stack pointer into the same stack (recursive locking), i.e.:
91 //
92 // 1) (hdr & aligned_mask) == 0
93 // 2) rsp <= hdr
94 // 3) hdr <= rsp + page_size
95 //
96 // these 3 tests can be done by evaluating the following expression:
97 //
98 // (hdr - rsp) & (aligned_mask - page_size)
99 //
152 jcc(Assembler::notEqual, slow_case);
153 // done
154 bind(done);
155 dec_held_monitor_count();
156 }
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, t1, t2);
172 #ifdef _LP64
173 if (UseCompactObjectHeaders) {
174 movptr(t1, Address(klass, Klass::prototype_header_offset()));
175 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), t1);
176 } else if (UseCompressedClassPointers) { // Take care not to kill klass
177 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), checked_cast<int32_t>(markWord::prototype().value()));
178 movptr(t1, klass);
179 encode_klass_not_null(t1, rscratch1);
180 movl(Address(obj, oopDesc::klass_offset_in_bytes()), t1);
181 } else
182 #endif
183 {
184 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), checked_cast<int32_t>(markWord::prototype().value()));
185 movptr(Address(obj, oopDesc::klass_offset_in_bytes()), klass);
186 }
187
188 if (len->is_valid()) {
189 movl(Address(obj, arrayOopDesc::length_offset_in_bytes()), len);
190 #ifdef _LP64
191 int base_offset = arrayOopDesc::length_offset_in_bytes() + BytesPerInt;
192 if (!is_aligned(base_offset, BytesPerWord)) {
193 assert(is_aligned(base_offset, BytesPerInt), "must be 4-byte aligned");
194 // Clear gap/first 4 bytes following the length field.
195 xorl(t1, t1);
196 movl(Address(obj, base_offset), t1);
197 }
198 #endif
199 }
200 #ifdef _LP64
201 else if (UseCompressedClassPointers && !UseCompactObjectHeaders) {
202 xorptr(t1, t1);
203 store_klass_gap(obj, t1);
204 }
302
303 initialize_header(obj, klass, len, t1, t2);
304
305 // clear rest of allocated space
306 if (zero_array) {
307 const Register len_zero = len;
308 // Align-up to word boundary, because we clear the 4 bytes potentially
309 // following the length field in initialize_header().
310 int base_offset = align_up(base_offset_in_bytes, BytesPerWord);
311 initialize_body(obj, arr_size, base_offset, len_zero);
312 }
313
314 if (CURRENT_ENV->dtrace_alloc_probes()) {
315 assert(obj == rax, "must be");
316 call(RuntimeAddress(Runtime1::entry_for(C1StubId::dtrace_object_alloc_id)));
317 }
318
319 verify_oop(obj);
320 }
321
322 void C1_MacroAssembler::build_frame(int frame_size_in_bytes, int bang_size_in_bytes) {
323 assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
324 // Make sure there is enough stack space for this method's activation.
325 // Note that we do this before doing an enter(). This matches the
326 // ordering of C2's stack overflow check / rsp decrement and allows
327 // the SharedRuntime stack overflow handling to be consistent
328 // between the two compilers.
329 generate_stack_overflow_check(bang_size_in_bytes);
330
331 push(rbp);
332 if (PreserveFramePointer) {
333 mov(rbp, rsp);
334 }
335 decrement(rsp, frame_size_in_bytes); // does not emit code for frame_size == 0
336
337 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
338 // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
339 bs->nmethod_entry_barrier(this, nullptr /* slow_path */, nullptr /* continuation */);
340 }
341
342
343 void C1_MacroAssembler::remove_frame(int frame_size_in_bytes) {
344 increment(rsp, frame_size_in_bytes); // Does not emit code for frame_size == 0
345 pop(rbp);
346 }
347
348
349 void C1_MacroAssembler::verified_entry(bool breakAtEntry) {
350 if (breakAtEntry || VerifyFPU) {
351 // Verified Entry first instruction should be 5 bytes long for correct
352 // patching by patch_verified_entry().
353 //
354 // Breakpoint and VerifyFPU have one byte first instruction.
355 // Also first instruction will be one byte "push(rbp)" if stack banging
356 // code is not generated (see build_frame() above).
357 // For all these cases generate long instruction first.
358 fat_nop();
359 }
360 if (breakAtEntry) int3();
361 // build frame
362 IA32_ONLY( verify_FPU(0, "method_entry"); )
363 }
364
365 void C1_MacroAssembler::load_parameter(int offset_in_words, Register reg) {
366 // rbp, + 0: link
367 // + 1: return address
368 // + 2: argument with offset 0
369 // + 3: argument with offset 1
370 // + 4: ...
371
372 movptr(reg, Address(rbp, (offset_in_words + 2) * BytesPerWord));
373 }
374
375 #ifndef PRODUCT
376
377 void C1_MacroAssembler::verify_stack_oop(int stack_offset) {
378 if (!VerifyOops) return;
379 verify_oop_addr(Address(rsp, stack_offset));
380 }
381
382 void C1_MacroAssembler::verify_not_null_oop(Register r) {
383 if (!VerifyOops) return;
384 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 "c1/c1_MacroAssembler.hpp"
26 #include "c1/c1_Runtime1.hpp"
27 #include "code/compiledIC.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/globals.hpp"
39 #include "runtime/os.hpp"
40 #include "runtime/sharedRuntime.hpp"
41 #include "runtime/stubRoutines.hpp"
42 #include "utilities/checkedCast.hpp"
43 #include "utilities/globalDefinitions.hpp"
44
45 int C1_MacroAssembler::lock_object(Register hdr, Register obj, Register disp_hdr, Register tmp, Label& slow_case) {
46 const int aligned_mask = BytesPerWord -1;
47 const int hdr_offset = oopDesc::mark_offset_in_bytes();
48 assert(hdr == rax, "hdr must be rax, for the cmpxchg instruction");
49 assert_different_registers(hdr, obj, disp_hdr, tmp);
50 int null_check_offset = -1;
51
52 verify_oop(obj);
53
54 // save object being locked into the BasicObjectLock
55 movptr(Address(disp_hdr, BasicObjectLock::obj_offset()), obj);
56
57 null_check_offset = offset();
61 testb(Address(hdr, Klass::misc_flags_offset()), KlassFlags::_misc_is_value_based_class);
62 jcc(Assembler::notZero, slow_case);
63 }
64
65 if (LockingMode == LM_LIGHTWEIGHT) {
66 #ifdef _LP64
67 const Register thread = r15_thread;
68 lightweight_lock(disp_hdr, obj, hdr, thread, tmp, slow_case);
69 #else
70 // Implicit null check.
71 movptr(hdr, Address(obj, oopDesc::mark_offset_in_bytes()));
72 // Lacking registers and thread on x86_32. Always take slow path.
73 jmp(slow_case);
74 #endif
75 } else if (LockingMode == LM_LEGACY) {
76 Label done;
77 // Load object header
78 movptr(hdr, Address(obj, hdr_offset));
79 // and mark it as unlocked
80 orptr(hdr, markWord::unlocked_value);
81 if (EnableValhalla) {
82 // Mask inline_type bit such that we go to the slow path if object is an inline type
83 andptr(hdr, ~((int) markWord::inline_type_bit_in_place));
84 }
85 // save unlocked object header into the displaced header location on the stack
86 movptr(Address(disp_hdr, 0), hdr);
87 // test if object header is still the same (i.e. unlocked), and if so, store the
88 // displaced header address in the object header - if it is not the same, get the
89 // object header instead
90 MacroAssembler::lock(); // must be immediately before cmpxchg!
91 cmpxchgptr(disp_hdr, Address(obj, hdr_offset));
92 // if the object header was the same, we're done
93 jcc(Assembler::equal, done);
94 // if the object header was not the same, it is now in the hdr register
95 // => test if it is a stack pointer into the same stack (recursive locking), i.e.:
96 //
97 // 1) (hdr & aligned_mask) == 0
98 // 2) rsp <= hdr
99 // 3) hdr <= rsp + page_size
100 //
101 // these 3 tests can be done by evaluating the following expression:
102 //
103 // (hdr - rsp) & (aligned_mask - page_size)
104 //
157 jcc(Assembler::notEqual, slow_case);
158 // done
159 bind(done);
160 dec_held_monitor_count();
161 }
162 }
163
164
165 // Defines obj, preserves var_size_in_bytes
166 void C1_MacroAssembler::try_allocate(Register obj, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2, Label& slow_case) {
167 if (UseTLAB) {
168 tlab_allocate(noreg, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
169 } else {
170 jmp(slow_case);
171 }
172 }
173
174
175 void C1_MacroAssembler::initialize_header(Register obj, Register klass, Register len, Register t1, Register t2) {
176 assert_different_registers(obj, klass, len, t1, t2);
177 if (UseCompactObjectHeaders || EnableValhalla) {
178 movptr(t1, Address(klass, Klass::prototype_header_offset()));
179 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), t1);
180 } else {
181 movptr(Address(obj, oopDesc::mark_offset_in_bytes()), checked_cast<int32_t>(markWord::prototype().value()));
182 }
183 #ifdef _LP64
184 if (UseCompressedClassPointers) { // Take care not to kill klass
185 movptr(t1, klass);
186 encode_klass_not_null(t1, rscratch1);
187 movl(Address(obj, oopDesc::klass_offset_in_bytes()), t1);
188 } else if (!UseCompactObjectHeaders)
189 #endif
190 {
191 movptr(Address(obj, oopDesc::klass_offset_in_bytes()), klass);
192 }
193
194 if (len->is_valid()) {
195 movl(Address(obj, arrayOopDesc::length_offset_in_bytes()), len);
196 #ifdef _LP64
197 int base_offset = arrayOopDesc::length_offset_in_bytes() + BytesPerInt;
198 if (!is_aligned(base_offset, BytesPerWord)) {
199 assert(is_aligned(base_offset, BytesPerInt), "must be 4-byte aligned");
200 // Clear gap/first 4 bytes following the length field.
201 xorl(t1, t1);
202 movl(Address(obj, base_offset), t1);
203 }
204 #endif
205 }
206 #ifdef _LP64
207 else if (UseCompressedClassPointers && !UseCompactObjectHeaders) {
208 xorptr(t1, t1);
209 store_klass_gap(obj, t1);
210 }
308
309 initialize_header(obj, klass, len, t1, t2);
310
311 // clear rest of allocated space
312 if (zero_array) {
313 const Register len_zero = len;
314 // Align-up to word boundary, because we clear the 4 bytes potentially
315 // following the length field in initialize_header().
316 int base_offset = align_up(base_offset_in_bytes, BytesPerWord);
317 initialize_body(obj, arr_size, base_offset, len_zero);
318 }
319
320 if (CURRENT_ENV->dtrace_alloc_probes()) {
321 assert(obj == rax, "must be");
322 call(RuntimeAddress(Runtime1::entry_for(C1StubId::dtrace_object_alloc_id)));
323 }
324
325 verify_oop(obj);
326 }
327
328 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) {
329 push(rbp);
330 if (PreserveFramePointer) {
331 mov(rbp, rsp);
332 }
333 decrement(rsp, frame_size_in_bytes);
334
335 if (needs_stack_repair) {
336 // Save stack increment (also account for fixed framesize and rbp)
337 assert((sp_inc & (StackAlignmentInBytes-1)) == 0, "stack increment not aligned");
338 int real_frame_size = sp_inc + frame_size_in_bytes + wordSize;
339 movptr(Address(rsp, frame_size_in_bytes - wordSize), real_frame_size);
340 }
341 if (reset_orig_pc) {
342 // Zero orig_pc to detect deoptimization during buffering in the entry points
343 movptr(Address(rsp, sp_offset_for_orig_pc), 0);
344 }
345 }
346
347 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) {
348 // Make sure there is enough stack space for this method's activation.
349 // Note that we do this before doing an enter(). This matches the
350 // ordering of C2's stack overflow check / rsp decrement and allows
351 // the SharedRuntime stack overflow handling to be consistent
352 // between the two compilers.
353 assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
354 generate_stack_overflow_check(bang_size_in_bytes);
355
356 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, 0, has_scalarized_args, needs_stack_repair);
357
358 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
359 // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
360 bs->nmethod_entry_barrier(this, nullptr /* slow_path */, nullptr /* continuation */);
361
362 if (verified_inline_entry_label != nullptr) {
363 // Jump here from the scalarized entry points that already created the frame.
364 bind(*verified_inline_entry_label);
365 }
366 }
367
368 void C1_MacroAssembler::verified_entry(bool breakAtEntry) {
369 if (breakAtEntry || VerifyFPU) {
370 // Verified Entry first instruction should be 5 bytes long for correct
371 // patching by patch_verified_entry().
372 //
373 // Breakpoint and VerifyFPU have one byte first instruction.
374 // Also first instruction will be one byte "push(rbp)" if stack banging
375 // code is not generated (see build_frame() above).
376 // For all these cases generate long instruction first.
377 fat_nop();
378 }
379 if (breakAtEntry) int3();
380 // build frame
381 IA32_ONLY( verify_FPU(0, "method_entry"); )
382 }
383
384 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) {
385 assert(InlineTypePassFieldsAsArgs, "sanity");
386 // Make sure there is enough stack space for this method's activation.
387 assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
388 generate_stack_overflow_check(bang_size_in_bytes);
389
390 GrowableArray<SigEntry>* sig = ces->sig();
391 GrowableArray<SigEntry>* sig_cc = is_inline_ro_entry ? ces->sig_cc_ro() : ces->sig_cc();
392 VMRegPair* regs = ces->regs();
393 VMRegPair* regs_cc = is_inline_ro_entry ? ces->regs_cc_ro() : ces->regs_cc();
394 int args_on_stack = ces->args_on_stack();
395 int args_on_stack_cc = is_inline_ro_entry ? ces->args_on_stack_cc_ro() : ces->args_on_stack_cc();
396
397 assert(sig->length() <= sig_cc->length(), "Zero-sized inline class not allowed!");
398 BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, sig_cc->length());
399 int args_passed = sig->length();
400 int args_passed_cc = SigEntry::fill_sig_bt(sig_cc, sig_bt);
401
402 // Create a temp frame so we can call into the runtime. It must be properly set up to accommodate GC.
403 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, 0, true, ces->c1_needs_stack_repair());
404
405 // The runtime call might safepoint, make sure nmethod entry barrier is executed
406 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
407 // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
408 bs->nmethod_entry_barrier(this, nullptr /* slow_path */, nullptr /* continuation */);
409
410 // FIXME -- call runtime only if we cannot in-line allocate all the incoming inline type args.
411 movptr(rbx, (intptr_t)(ces->method()));
412 if (is_inline_ro_entry) {
413 call(RuntimeAddress(Runtime1::entry_for(C1StubId::buffer_inline_args_no_receiver_id)));
414 } else {
415 call(RuntimeAddress(Runtime1::entry_for(C1StubId::buffer_inline_args_id)));
416 }
417 int rt_call_offset = offset();
418
419 // Remove the temp frame
420 addptr(rsp, frame_size_in_bytes);
421 pop(rbp);
422
423 // Check if we need to extend the stack for packing
424 int sp_inc = 0;
425 if (args_on_stack > args_on_stack_cc) {
426 sp_inc = extend_stack_for_inline_args(args_on_stack);
427 }
428
429 shuffle_inline_args(true, is_inline_ro_entry, sig_cc,
430 args_passed_cc, args_on_stack_cc, regs_cc, // from
431 args_passed, args_on_stack, regs, // to
432 sp_inc, rax);
433
434 // Create the real frame. Below jump will then skip over the stack banging and frame
435 // setup code in the verified_inline_entry (which has a different real_frame_size).
436 build_frame_helper(frame_size_in_bytes, sp_offset_for_orig_pc, sp_inc, false, ces->c1_needs_stack_repair());
437
438 jmp(verified_inline_entry_label);
439 return rt_call_offset;
440 }
441
442 void C1_MacroAssembler::load_parameter(int offset_in_words, Register reg) {
443 // rbp, + 0: link
444 // + 1: return address
445 // + 2: argument with offset 0
446 // + 3: argument with offset 1
447 // + 4: ...
448
449 movptr(reg, Address(rbp, (offset_in_words + 2) * BytesPerWord));
450 }
451
452 #ifndef PRODUCT
453
454 void C1_MacroAssembler::verify_stack_oop(int stack_offset) {
455 if (!VerifyOops) return;
456 verify_oop_addr(Address(rsp, stack_offset));
457 }
458
459 void C1_MacroAssembler::verify_not_null_oop(Register r) {
460 if (!VerifyOops) return;
461 Label not_null;
|