< prev index next >

src/hotspot/cpu/x86/c1_MacroAssembler_x86.cpp

Print this page
*** 36,16 ***
  #include "runtime/basicLock.hpp"
  #include "runtime/os.hpp"
  #include "runtime/sharedRuntime.hpp"
  #include "runtime/stubRoutines.hpp"
  
! int C1_MacroAssembler::lock_object(Register hdr, Register obj, Register disp_hdr, Label& slow_case) {
    const int aligned_mask = BytesPerWord -1;
    const int hdr_offset = oopDesc::mark_offset_in_bytes();
    assert(hdr == rax, "hdr must be rax, for the cmpxchg instruction");
!   assert(hdr != obj && hdr != disp_hdr && obj != disp_hdr, "registers must be different");
-   Label done;
    int null_check_offset = -1;
  
    verify_oop(obj);
  
    // save object being locked into the BasicObjectLock
--- 36,15 ---
  #include "runtime/basicLock.hpp"
  #include "runtime/os.hpp"
  #include "runtime/sharedRuntime.hpp"
  #include "runtime/stubRoutines.hpp"
  
! int C1_MacroAssembler::lock_object(Register hdr, Register obj, Register disp_hdr, Register tmp, Label& slow_case) {
    const int aligned_mask = BytesPerWord -1;
    const int hdr_offset = oopDesc::mark_offset_in_bytes();
    assert(hdr == rax, "hdr must be rax, for the cmpxchg instruction");
!   assert_different_registers(hdr, obj, disp_hdr, tmp);
    int null_check_offset = -1;
  
    verify_oop(obj);
  
    // save object being locked into the BasicObjectLock

*** 60,43 ***
      jcc(Assembler::notZero, slow_case);
    }
  
    // Load object header
    movptr(hdr, Address(obj, hdr_offset));
!   // and mark it as unlocked
!   orptr(hdr, markWord::unlocked_value);
!   // save unlocked object header into the displaced header location on the stack
!   movptr(Address(disp_hdr, 0), hdr);
!   // test if object header is still the same (i.e. unlocked), and if so, store the
!   // displaced header address in the object header - if it is not the same, get the
!   // object header instead
!   MacroAssembler::lock(); // must be immediately before cmpxchg!
!   cmpxchgptr(disp_hdr, Address(obj, hdr_offset));
!   // if the object header was the same, we're done
!   jcc(Assembler::equal, done);
!   // if the object header was not the same, it is now in the hdr register
!   // => test if it is a stack pointer into the same stack (recursive locking), i.e.:
!   //
!   // 1) (hdr & aligned_mask) == 0
!   // 2) rsp <= hdr
!   // 3) hdr <= rsp + page_size
!   //
!   // these 3 tests can be done by evaluating the following expression:
!   //
!   // (hdr - rsp) & (aligned_mask - page_size)
!   //
!   // assuming both the stack pointer and page_size have their least
!   // significant 2 bits cleared and page_size is a power of 2
!   subptr(hdr, rsp);
!   andptr(hdr, aligned_mask - (int)os::vm_page_size());
!   // for recursive locking, the result is zero => save it in the displaced header
!   // location (NULL in the displaced hdr location indicates recursive locking)
!   movptr(Address(disp_hdr, 0), hdr);
!   // otherwise we don't care about the result and handle locking via runtime call
!   jcc(Assembler::notZero, slow_case);
!   // done
!   bind(done);
  
    inc_held_monitor_count();
  
    return null_check_offset;
  }
--- 59,54 ---
      jcc(Assembler::notZero, slow_case);
    }
  
    // Load object header
    movptr(hdr, Address(obj, hdr_offset));
! 
!   if (UseFastLocking) {
! #ifdef _LP64
!     const Register thread = r15_thread;
! #else
!     const Register thread = disp_hdr;
!     get_thread(thread);
! #endif
!     fast_lock_impl(obj, hdr, thread, tmp, slow_case, LP64_ONLY(false) NOT_LP64(true));
!   } else {
!     Label done;
!     orptr(hdr, markWord::unlocked_value);
!     // save unlocked object header into the displaced header location on the stack
!     movptr(Address(disp_hdr, 0), hdr);
!     // test if object header is still the same (i.e. unlocked), and if so, store the
!     // displaced header address in the object header - if it is not the same, get the
!     // object header instead
!     MacroAssembler::lock(); // must be immediately before cmpxchg!
!     cmpxchgptr(disp_hdr, Address(obj, hdr_offset));
!     // if the object header was the same, we're done
!     jcc(Assembler::equal, done);
!     // if the object header was not the same, it is now in the hdr register
!     // => test if it is a stack pointer into the same stack (recursive locking), i.e.:
!     //
!     // 1) (hdr & aligned_mask) == 0
!     // 2) rsp <= hdr
!     // 3) hdr <= rsp + page_size
!     //
!     // these 3 tests can be done by evaluating the following expression:
!     //
!     // (hdr - rsp) & (aligned_mask - page_size)
!     //
!     // assuming both the stack pointer and page_size have their least
+     // significant 2 bits cleared and page_size is a power of 2
+     subptr(hdr, rsp);
+     andptr(hdr, aligned_mask - (int)os::vm_page_size());
+     // for recursive locking, the result is zero => save it in the displaced header
+     // location (NULL in the displaced hdr location indicates recursive locking)
+     movptr(Address(disp_hdr, 0), hdr);
+     // otherwise we don't care about the result and handle locking via runtime call
+     jcc(Assembler::notZero, slow_case);
+     // done
+     bind(done);
+   }
  
    inc_held_monitor_count();
  
    return null_check_offset;
  }

*** 106,31 ***
    const int hdr_offset = oopDesc::mark_offset_in_bytes();
    assert(disp_hdr == rax, "disp_hdr must be rax, for the cmpxchg instruction");
    assert(hdr != obj && hdr != disp_hdr && obj != disp_hdr, "registers must be different");
    Label done;
  
!   // load displaced header
!   movptr(hdr, Address(disp_hdr, 0));
!   // if the loaded hdr is NULL we had recursive locking
!   testptr(hdr, hdr);
!   // if we had recursive locking, we are done
!   jcc(Assembler::zero, done);
    // load object
    movptr(obj, Address(disp_hdr, BasicObjectLock::obj_offset_in_bytes()));
- 
    verify_oop(obj);
-   // test if object header is pointing to the displaced header, and if so, restore
-   // the displaced header in the object - if the object header is not pointing to
-   // the displaced header, get the object header instead
-   MacroAssembler::lock(); // must be immediately before cmpxchg!
-   cmpxchgptr(hdr, Address(obj, hdr_offset));
-   // if the object header was not pointing to the displaced header,
-   // we do unlocking via runtime call
-   jcc(Assembler::notEqual, slow_case);
-   // done
-   bind(done);
  
    dec_held_monitor_count();
  }
  
  
  // Defines obj, preserves var_size_in_bytes
--- 116,39 ---
    const int hdr_offset = oopDesc::mark_offset_in_bytes();
    assert(disp_hdr == rax, "disp_hdr must be rax, for the cmpxchg instruction");
    assert(hdr != obj && hdr != disp_hdr && obj != disp_hdr, "registers must be different");
    Label done;
  
!   if (!UseFastLocking) {
!     // load displaced header
!     movptr(hdr, Address(disp_hdr, 0));
!     // if the loaded hdr is NULL we had recursive locking
!     testptr(hdr, hdr);
!     // if we had recursive locking, we are done
+     jcc(Assembler::zero, done);
+   }
+ 
    // load object
    movptr(obj, Address(disp_hdr, BasicObjectLock::obj_offset_in_bytes()));
    verify_oop(obj);
  
+   if (UseFastLocking) {
+     movptr(disp_hdr, Address(obj, hdr_offset));
+     andptr(disp_hdr, ~(int32_t)markWord::lock_mask_in_place);
+     fast_unlock_impl(obj, disp_hdr, hdr, slow_case);
+   } else {
+     // test if object header is pointing to the displaced header, and if so, restore
+     // the displaced header in the object - if the object header is not pointing to
+     // the displaced header, get the object header instead
+     MacroAssembler::lock(); // must be immediately before cmpxchg!
+     cmpxchgptr(hdr, Address(obj, hdr_offset));
+     // if the object header was not pointing to the displaced header,
+     // we do unlocking via runtime call
+     jcc(Assembler::notEqual, slow_case);
+     // done
+   }
+   bind(done);
    dec_held_monitor_count();
  }
  
  
  // Defines obj, preserves var_size_in_bytes

*** 142,28 ***
    }
  }
  
  
  void C1_MacroAssembler::initialize_header(Register obj, Register klass, Register len, Register t1, Register t2) {
!   assert_different_registers(obj, klass, len);
!   movptr(Address(obj, oopDesc::mark_offset_in_bytes()), checked_cast<int32_t>(markWord::prototype().value()));
  #ifdef _LP64
!   if (UseCompressedClassPointers) { // Take care not to kill klass
!     movptr(t1, klass);
!     encode_klass_not_null(t1, rscratch1);
!     movl(Address(obj, oopDesc::klass_offset_in_bytes()), t1);
!   } else
  #endif
!   {
!     movptr(Address(obj, oopDesc::klass_offset_in_bytes()), klass);
    }
- 
    if (len->is_valid()) {
      movl(Address(obj, arrayOopDesc::length_offset_in_bytes()), len);
    }
  #ifdef _LP64
!   else if (UseCompressedClassPointers) {
      xorptr(t1, t1);
      store_klass_gap(obj, t1);
    }
  #endif
  }
--- 160,32 ---
    }
  }
  
  
  void C1_MacroAssembler::initialize_header(Register obj, Register klass, Register len, Register t1, Register t2) {
!   assert_different_registers(obj, klass, len, t1, t2);
!   if (UseCompactObjectHeaders) {
+     movptr(t1, Address(klass, Klass::prototype_header_offset()));
+     movptr(Address(obj, oopDesc::mark_offset_in_bytes()), t1);
+   } else {
+     movptr(Address(obj, oopDesc::mark_offset_in_bytes()), checked_cast<int32_t>(markWord::prototype().value()));
  #ifdef _LP64
!     if (UseCompressedClassPointers) { // Take care not to kill klass
!       movptr(t1, klass);
!       encode_klass_not_null(t1, rscratch1);
!       movl(Address(obj, oopDesc::klass_offset_in_bytes()), t1);
!     } else
  #endif
!     {
!       movptr(Address(obj, oopDesc::klass_offset_in_bytes()), klass);
+     }
    }
    if (len->is_valid()) {
      movl(Address(obj, arrayOopDesc::length_offset_in_bytes()), len);
    }
  #ifdef _LP64
!   else if (UseCompressedClassPointers && !UseCompactObjectHeaders) {
      xorptr(t1, t1);
      store_klass_gap(obj, t1);
    }
  #endif
  }

*** 201,34 ***
    if (!(UseTLAB && ZeroTLAB && is_tlab_allocated)) {
      // clear rest of allocated space
      const Register t1_zero = t1;
      const Register index = t2;
      const int threshold = 6 * BytesPerWord;   // approximate break even point for code size (see comments below)
      if (var_size_in_bytes != noreg) {
        mov(index, var_size_in_bytes);
!       initialize_body(obj, index, hdr_size_in_bytes, t1_zero);
      } else if (con_size_in_bytes <= threshold) {
        // use explicit null stores
        // code size = 2 + 3*n bytes (n = number of fields to clear)
        xorptr(t1_zero, t1_zero); // use t1_zero reg to clear memory (shorter code)
!       for (int i = hdr_size_in_bytes; i < con_size_in_bytes; i += BytesPerWord)
          movptr(Address(obj, i), t1_zero);
!     } else if (con_size_in_bytes > hdr_size_in_bytes) {
        // use loop to null out the fields
        // code size = 16 bytes for even n (n = number of fields to clear)
        // initialize last object field first if odd number of fields
        xorptr(t1_zero, t1_zero); // use t1_zero reg to clear memory (shorter code)
!       movptr(index, (con_size_in_bytes - hdr_size_in_bytes) >> 3);
        // initialize last object field if constant size is odd
!       if (((con_size_in_bytes - hdr_size_in_bytes) & 4) != 0)
          movptr(Address(obj, con_size_in_bytes - (1*BytesPerWord)), t1_zero);
        // initialize remaining object fields: rdx is a multiple of 2
        { Label loop;
          bind(loop);
!         movptr(Address(obj, index, Address::times_8, hdr_size_in_bytes - (1*BytesPerWord)),
                 t1_zero);
!         NOT_LP64(movptr(Address(obj, index, Address::times_8, hdr_size_in_bytes - (2*BytesPerWord)),
                 t1_zero);)
          decrement(index);
          jcc(Assembler::notZero, loop);
        }
      }
--- 223,35 ---
    if (!(UseTLAB && ZeroTLAB && is_tlab_allocated)) {
      // clear rest of allocated space
      const Register t1_zero = t1;
      const Register index = t2;
      const int threshold = 6 * BytesPerWord;   // approximate break even point for code size (see comments below)
+     int hdr_size_aligned = align_up(hdr_size_in_bytes, BytesPerWord); // klass gap is already cleared by init_header().
      if (var_size_in_bytes != noreg) {
        mov(index, var_size_in_bytes);
!       initialize_body(obj, index, hdr_size_aligned, t1_zero);
      } else if (con_size_in_bytes <= threshold) {
        // use explicit null stores
        // code size = 2 + 3*n bytes (n = number of fields to clear)
        xorptr(t1_zero, t1_zero); // use t1_zero reg to clear memory (shorter code)
!       for (int i = hdr_size_aligned; i < con_size_in_bytes; i += BytesPerWord)
          movptr(Address(obj, i), t1_zero);
!     } else if (con_size_in_bytes > hdr_size_aligned) {
        // use loop to null out the fields
        // code size = 16 bytes for even n (n = number of fields to clear)
        // initialize last object field first if odd number of fields
        xorptr(t1_zero, t1_zero); // use t1_zero reg to clear memory (shorter code)
!       movptr(index, (con_size_in_bytes - hdr_size_aligned) >> 3);
        // initialize last object field if constant size is odd
!       if (((con_size_in_bytes - hdr_size_aligned) & 4) != 0)
          movptr(Address(obj, con_size_in_bytes - (1*BytesPerWord)), t1_zero);
        // initialize remaining object fields: rdx is a multiple of 2
        { Label loop;
          bind(loop);
!         movptr(Address(obj, index, Address::times_8, hdr_size_aligned - (1*BytesPerWord)),
                 t1_zero);
!         NOT_LP64(movptr(Address(obj, index, Address::times_8, hdr_size_aligned - (2*BytesPerWord)),
                 t1_zero);)
          decrement(index);
          jcc(Assembler::notZero, loop);
        }
      }

*** 240,11 ***
    }
  
    verify_oop(obj);
  }
  
! void C1_MacroAssembler::allocate_array(Register obj, Register len, Register t1, Register t2, int header_size, Address::ScaleFactor f, Register klass, Label& slow_case) {
    assert(obj == rax, "obj must be in rax, for cmpxchg");
    assert_different_registers(obj, len, t1, t2, klass);
  
    // determine alignment mask
    assert(!(BytesPerWord & 1), "must be a multiple of 2 for masking code to work");
--- 263,11 ---
    }
  
    verify_oop(obj);
  }
  
! 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) {
    assert(obj == rax, "obj must be in rax, for cmpxchg");
    assert_different_registers(obj, len, t1, t2, klass);
  
    // determine alignment mask
    assert(!(BytesPerWord & 1), "must be a multiple of 2 for masking code to work");

*** 253,21 ***
    cmpptr(len, checked_cast<int32_t>(max_array_allocation_length));
    jcc(Assembler::above, slow_case);
  
    const Register arr_size = t2; // okay to be the same
    // align object end
!   movptr(arr_size, header_size * BytesPerWord + MinObjAlignmentInBytesMask);
    lea(arr_size, Address(arr_size, len, f));
    andptr(arr_size, ~MinObjAlignmentInBytesMask);
  
    try_allocate(obj, arr_size, 0, t1, t2, slow_case);
  
    initialize_header(obj, klass, len, t1, t2);
  
    // clear rest of allocated space
    const Register len_zero = len;
!   initialize_body(obj, arr_size, header_size * BytesPerWord, len_zero);
  
    if (CURRENT_ENV->dtrace_alloc_probes()) {
      assert(obj == rax, "must be");
      call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id)));
    }
--- 276,21 ---
    cmpptr(len, checked_cast<int32_t>(max_array_allocation_length));
    jcc(Assembler::above, slow_case);
  
    const Register arr_size = t2; // okay to be the same
    // align object end
!   movptr(arr_size, (int32_t)base_offset_in_bytes + MinObjAlignmentInBytesMask);
    lea(arr_size, Address(arr_size, len, f));
    andptr(arr_size, ~MinObjAlignmentInBytesMask);
  
    try_allocate(obj, arr_size, 0, t1, t2, slow_case);
  
    initialize_header(obj, klass, len, t1, t2);
  
    // clear rest of allocated space
    const Register len_zero = len;
!   initialize_body(obj, arr_size, base_offset_in_bytes, len_zero);
  
    if (CURRENT_ENV->dtrace_alloc_probes()) {
      assert(obj == rax, "must be");
      call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id)));
    }

*** 297,11 ***
    const int ic_cmp_size = LP64_ONLY(10) NOT_LP64(9);
    assert(UseCompressedClassPointers || offset() - start_offset == ic_cmp_size, "check alignment in emit_method_entry");
  }
  
  
! void C1_MacroAssembler::build_frame(int frame_size_in_bytes, int bang_size_in_bytes) {
    assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
    // Make sure there is enough stack space for this method's activation.
    // Note that we do this before doing an enter(). This matches the
    // ordering of C2's stack overflow check / rsp decrement and allows
    // the SharedRuntime stack overflow handling to be consistent
--- 320,11 ---
    const int ic_cmp_size = LP64_ONLY(10) NOT_LP64(9);
    assert(UseCompressedClassPointers || offset() - start_offset == ic_cmp_size, "check alignment in emit_method_entry");
  }
  
  
! void C1_MacroAssembler::build_frame(int frame_size_in_bytes, int bang_size_in_bytes, int max_monitors) {
    assert(bang_size_in_bytes >= frame_size_in_bytes, "stack bang size incorrect");
    // Make sure there is enough stack space for this method's activation.
    // Note that we do this before doing an enter(). This matches the
    // ordering of C2's stack overflow check / rsp decrement and allows
    // the SharedRuntime stack overflow handling to be consistent

*** 318,10 ***
--- 341,23 ---
      empty_FPU_stack();
    }
  #endif // !_LP64 && COMPILER2
    decrement(rsp, frame_size_in_bytes); // does not emit code for frame_size == 0
  
+ #ifdef _LP64
+   if (UseFastLocking && max_monitors > 0) {
+     Label ok;
+     movptr(rax, Address(r15_thread, JavaThread::lock_stack_current_offset()));
+     addptr(rax, max_monitors * wordSize);
+     cmpptr(rax, Address(r15_thread, JavaThread::lock_stack_limit_offset()));
+     jcc(Assembler::less, ok);
+     assert(StubRoutines::x86::check_lock_stack() != NULL, "need runtime call stub");
+     call(RuntimeAddress(StubRoutines::x86::check_lock_stack()));
+     bind(ok);
+   }
+ #endif
+ 
    BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
    // C1 code is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
    bs->nmethod_entry_barrier(this, NULL /* slow_path */, NULL /* continuation */);
  }
  
< prev index next >