< prev index next >

src/hotspot/cpu/x86/c1_MacroAssembler_x86.cpp

Print this page

 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, 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(hdr != obj && hdr != disp_hdr && obj != disp_hdr, "registers must be different");
 46   Label done;
 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_in_bytes()), 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   // and mark it as unlocked
 66   orptr(hdr, markWord::unlocked_value);




 67   // save unlocked object header into the displaced header location on the stack
 68   movptr(Address(disp_hdr, 0), hdr);
 69   // test if object header is still the same (i.e. unlocked), and if so, store the
 70   // displaced header address in the object header - if it is not the same, get the
 71   // object header instead
 72   MacroAssembler::lock(); // must be immediately before cmpxchg!
 73   cmpxchgptr(disp_hdr, Address(obj, hdr_offset));
 74   // if the object header was the same, we're done
 75   jcc(Assembler::equal, done);
 76   // if the object header was not the same, it is now in the hdr register
 77   // => test if it is a stack pointer into the same stack (recursive locking), i.e.:
 78   //
 79   // 1) (hdr & aligned_mask) == 0
 80   // 2) rsp <= hdr
 81   // 3) hdr <= rsp + page_size
 82   //
 83   // these 3 tests can be done by evaluating the following expression:
 84   //
 85   // (hdr - rsp) & (aligned_mask - page_size)
 86   //

128   jcc(Assembler::notEqual, slow_case);
129   // done
130   bind(done);
131 
132   dec_held_monitor_count();
133 }
134 
135 
136 // Defines obj, preserves var_size_in_bytes
137 void C1_MacroAssembler::try_allocate(Register obj, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2, Label& slow_case) {
138   if (UseTLAB) {
139     tlab_allocate(noreg, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
140   } else {
141     jmp(slow_case);
142   }
143 }
144 
145 
146 void C1_MacroAssembler::initialize_header(Register obj, Register klass, Register len, Register t1, Register t2) {
147   assert_different_registers(obj, klass, len);
148   movptr(Address(obj, oopDesc::mark_offset_in_bytes()), checked_cast<int32_t>(markWord::prototype().value()));








149 #ifdef _LP64
150   if (UseCompressedClassPointers) { // Take care not to kill klass
151     movptr(t1, klass);
152     encode_klass_not_null(t1, rscratch1);
153     movl(Address(obj, oopDesc::klass_offset_in_bytes()), t1);
154   } else
155 #endif
156   {
157     movptr(Address(obj, oopDesc::klass_offset_in_bytes()), klass);
158   }
159 
160   if (len->is_valid()) {
161     movl(Address(obj, arrayOopDesc::length_offset_in_bytes()), len);
162   }
163 #ifdef _LP64
164   else if (UseCompressedClassPointers) {
165     xorptr(t1, t1);
166     store_klass_gap(obj, t1);
167   }
168 #endif

282   verify_oop(receiver);
283   // explicit NULL check not needed since load from [klass_offset] causes a trap
284   // check against inline cache
285   assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()), "must add explicit null check");
286   int start_offset = offset();
287 
288   if (UseCompressedClassPointers) {
289     load_klass(rscratch1, receiver, rscratch2);
290     cmpptr(rscratch1, iCache);
291   } else {
292     cmpptr(iCache, Address(receiver, oopDesc::klass_offset_in_bytes()));
293   }
294   // if icache check fails, then jump to runtime routine
295   // Note: RECEIVER must still contain the receiver!
296   jump_cc(Assembler::notEqual,
297           RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
298   const int ic_cmp_size = LP64_ONLY(10) NOT_LP64(9);
299   assert(UseCompressedClassPointers || offset() - start_offset == ic_cmp_size, "check alignment in emit_method_entry");
300 }
301 












302 
303 void C1_MacroAssembler::build_frame(int frame_size_in_bytes, int bang_size_in_bytes) {
304   assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");














305   // Make sure there is enough stack space for this method's activation.
306   // Note that we do this before doing an enter(). This matches the
307   // ordering of C2's stack overflow check / rsp decrement and allows
308   // the SharedRuntime stack overflow handling to be consistent
309   // between the two compilers.

310   generate_stack_overflow_check(bang_size_in_bytes);
311 
312   push(rbp);
313   if (PreserveFramePointer) {
314     mov(rbp, rsp);
315   }
316 #if !defined(_LP64) && defined(COMPILER2)
317   if (UseSSE < 2 && !CompilerConfig::is_c1_only_no_jvmci()) {
318     // c2 leaves fpu stack dirty. Clean it on entry
319     empty_FPU_stack();
320   }
321 #endif // !_LP64 && COMPILER2
322   decrement(rsp, frame_size_in_bytes); // does not emit code for frame_size == 0
323 
324   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
325   // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
326   bs->nmethod_entry_barrier(this, NULL /* slow_path */, NULL /* continuation */);
327 }
328 
329 
330 void C1_MacroAssembler::remove_frame(int frame_size_in_bytes) {
331   increment(rsp, frame_size_in_bytes);  // Does not emit code for frame_size == 0
332   pop(rbp);


333 }
334 
335 
336 void C1_MacroAssembler::verified_entry(bool breakAtEntry) {
337   if (breakAtEntry || VerifyFPU) {
338     // Verified Entry first instruction should be 5 bytes long for correct
339     // patching by patch_verified_entry().
340     //
341     // Breakpoint and VerifyFPU have one byte first instruction.
342     // Also first instruction will be one byte "push(rbp)" if stack banging
343     // code is not generated (see build_frame() above).
344     // For all these cases generate long instruction first.
345     fat_nop();
346   }
347   if (breakAtEntry) int3();
348   // build frame
349   IA32_ONLY( verify_FPU(0, "method_entry"); )
350 }
351 






























































352 void C1_MacroAssembler::load_parameter(int offset_in_words, Register reg) {
353   // rbp, + 0: link
354   //     + 1: return address
355   //     + 2: argument with offset 0
356   //     + 3: argument with offset 1
357   //     + 4: ...
358 
359   movptr(reg, Address(rbp, (offset_in_words + 2) * BytesPerWord));
360 }
361 
362 #ifndef PRODUCT
363 
364 void C1_MacroAssembler::verify_stack_oop(int stack_offset) {
365   if (!VerifyOops) return;
366   verify_oop_addr(Address(rsp, stack_offset));
367 }
368 
369 void C1_MacroAssembler::verify_not_null_oop(Register r) {
370   if (!VerifyOops) return;
371   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, 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   //

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

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_inc, 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 }
335 
336 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) {
337   if (has_scalarized_args) {
338     // Initialize orig_pc to detect deoptimization during buffering in the entry points
339     movptr(Address(rsp, sp_offset_for_orig_pc - frame_size_in_bytes - wordSize), 0);
340   }
341   if (!needs_stack_repair && verified_inline_entry_label != NULL) {
342     bind(*verified_inline_entry_label);
343   }
344   // Make sure there is enough stack space for this method's activation.
345   // Note that we do this before doing an enter(). This matches the
346   // ordering of C2's stack overflow check / rsp decrement and allows
347   // the SharedRuntime stack overflow handling to be consistent
348   // between the two compilers.
349   assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
350   generate_stack_overflow_check(bang_size_in_bytes);
351 
352   build_frame_helper(frame_size_in_bytes, 0, needs_stack_repair);










353 
354   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
355   // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
356   bs->nmethod_entry_barrier(this, NULL /* slow_path */, NULL /* continuation */);


357 
358   if (needs_stack_repair && verified_inline_entry_label != NULL) {
359     // Jump here from the scalarized entry points that require additional stack space
360     // for packing scalarized arguments and therefore already created the frame.
361     bind(*verified_inline_entry_label);
362   }
363 }
364 

365 void C1_MacroAssembler::verified_entry(bool breakAtEntry) {
366   if (breakAtEntry || VerifyFPU) {
367     // Verified Entry first instruction should be 5 bytes long for correct
368     // patching by patch_verified_entry().
369     //
370     // Breakpoint and VerifyFPU have one byte first instruction.
371     // Also first instruction will be one byte "push(rbp)" if stack banging
372     // code is not generated (see build_frame() above).
373     // For all these cases generate long instruction first.
374     fat_nop();
375   }
376   if (breakAtEntry) int3();
377   // build frame
378   IA32_ONLY( verify_FPU(0, "method_entry"); )
379 }
380 
381 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) {
382   assert(InlineTypePassFieldsAsArgs, "sanity");
383   // Make sure there is enough stack space for this method's activation.
384   assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
385   generate_stack_overflow_check(bang_size_in_bytes);
386 
387   GrowableArray<SigEntry>* sig    = &ces->sig();
388   GrowableArray<SigEntry>* sig_cc = is_inline_ro_entry ? &ces->sig_cc_ro() : &ces->sig_cc();
389   VMRegPair* regs      = ces->regs();
390   VMRegPair* regs_cc   = is_inline_ro_entry ? ces->regs_cc_ro() : ces->regs_cc();
391   int args_on_stack    = ces->args_on_stack();
392   int args_on_stack_cc = is_inline_ro_entry ? ces->args_on_stack_cc_ro() : ces->args_on_stack_cc();
393 
394   assert(sig->length() <= sig_cc->length(), "Zero-sized inline class not allowed!");
395   BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, sig_cc->length());
396   int args_passed = sig->length();
397   int args_passed_cc = SigEntry::fill_sig_bt(sig_cc, sig_bt);
398 
399   // Check if we need to extend the stack for packing
400   int sp_inc = 0;
401   if (args_on_stack > args_on_stack_cc) {
402     sp_inc = extend_stack_for_inline_args(args_on_stack);
403   }
404 
405   // Create a temp frame so we can call into the runtime. It must be properly set up to accommodate GC.
406   build_frame_helper(frame_size_in_bytes, sp_inc, ces->c1_needs_stack_repair());
407 
408   // Initialize orig_pc to detect deoptimization during buffering in below runtime call
409   movptr(Address(rsp, sp_offset_for_orig_pc), 0);
410 
411   // The runtime call might safepoint, make sure nmethod entry barrier is executed
412   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
413   bs->nmethod_entry_barrier(this, NULL /* slow_path */, NULL /* continuation */);
414 
415   // FIXME -- call runtime only if we cannot in-line allocate all the incoming inline type args.
416   movptr(rbx, (intptr_t)(ces->method()));
417   if (is_inline_ro_entry) {
418     call(RuntimeAddress(Runtime1::entry_for(Runtime1::buffer_inline_args_no_receiver_id)));
419   } else {
420     call(RuntimeAddress(Runtime1::entry_for(Runtime1::buffer_inline_args_id)));
421   }
422   int rt_call_offset = offset();
423 
424   // Remove the temp frame
425   addptr(rsp, frame_size_in_bytes);
426   pop(rbp);
427 
428   shuffle_inline_args(true, is_inline_ro_entry, sig_cc,
429                       args_passed_cc, args_on_stack_cc, regs_cc, // from
430                       args_passed, args_on_stack, regs,          // to
431                       sp_inc, rax);
432 
433   if (ces->c1_needs_stack_repair()) {
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_inc, true);
437   }
438 
439   jmp(verified_inline_entry_label);
440   return rt_call_offset;
441 }
442 
443 void C1_MacroAssembler::load_parameter(int offset_in_words, Register reg) {
444   // rbp, + 0: link
445   //     + 1: return address
446   //     + 2: argument with offset 0
447   //     + 3: argument with offset 1
448   //     + 4: ...
449 
450   movptr(reg, Address(rbp, (offset_in_words + 2) * BytesPerWord));
451 }
452 
453 #ifndef PRODUCT
454 
455 void C1_MacroAssembler::verify_stack_oop(int stack_offset) {
456   if (!VerifyOops) return;
457   verify_oop_addr(Address(rsp, stack_offset));
458 }
459 
460 void C1_MacroAssembler::verify_not_null_oop(Register r) {
461   if (!VerifyOops) return;
462   Label not_null;
< prev index next >