Reverting r4685, r4686, r4687

src/ia32/macro-assembler-ia32.cc

 // Copyright 2006-2009 the V8 project authors. All rights reserved.
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
 //
 //     * Redistributions of source code must retain the above copyright
 //       notice, this list of conditions and the following disclaimer.
 //     * Redistributions in binary form must reproduce the above
 //       copyright notice, this list of conditions and the following
 //       disclaimer in the documentation and/or other materials provided
@@ skipping 42 matching lines @@
                                        Register addr,
                                        Register scratch) {
   if (FLAG_debug_code) {
     // Check that the object is not in new space.
     Label not_in_new_space;
     InNewSpace(object, scratch, not_equal, &not_in_new_space);
     Abort("new-space object passed to RecordWriteHelper");
     bind(&not_in_new_space);
   }
 
+  Label fast;
+
   // Compute the page start address from the heap object pointer, and reuse
   // the 'object' register for it.
   and_(object, ~Page::kPageAlignmentMask);
+  Register page_start = object;
 
-  // Compute number of region covering addr. See Page::GetRegionNumberForAddress
-  // method for more details.
-  and_(addr, Page::kPageAlignmentMask);
-  shr(addr, Page::kRegionSizeLog2);
+  // Compute the bit addr in the remembered set/index of the pointer in the
+  // page. Reuse 'addr' as pointer_offset.
+  sub(addr, Operand(page_start));
+  shr(addr, kObjectAlignmentBits);
+  Register pointer_offset = addr;
 
-  // Set dirty mark for region.
-  bts(Operand(object, Page::kDirtyFlagOffset), addr);
+  // If the bit offset lies beyond the normal remembered set range, it is in
+  // the extra remembered set area of a large object.
+  cmp(pointer_offset, Page::kPageSize / kPointerSize);
+  j(less, &fast);
+
+  // Adjust 'page_start' so that addressing using 'pointer_offset' hits the
+  // extra remembered set after the large object.
+
+  // Find the length of the large object (FixedArray).
+  mov(scratch, Operand(page_start, Page::kObjectStartOffset
+                                         + FixedArray::kLengthOffset));
+  Register array_length = scratch;
+
+  // Extra remembered set starts right after the large object (a FixedArray), at
+  //   page_start + kObjectStartOffset + objectSize
+  // where objectSize is FixedArray::kHeaderSize + kPointerSize * array_length.
+  // Add the delta between the end of the normal RSet and the start of the
+  // extra RSet to 'page_start', so that addressing the bit using
+  // 'pointer_offset' hits the extra RSet words.
+  lea(page_start,
+      Operand(page_start, array_length, times_pointer_size,
+              Page::kObjectStartOffset + FixedArray::kHeaderSize
+                  - Page::kRSetEndOffset));
+
+  // NOTE: For now, we use the bit-test-and-set (bts) x86 instruction
+  // to limit code size. We should probably evaluate this decision by
+  // measuring the performance of an equivalent implementation using
+  // "simpler" instructions
+  bind(&fast);
+  bts(Operand(page_start, Page::kRSetOffset), pointer_offset);
 }
 
 
 void MacroAssembler::InNewSpace(Register object,
                                 Register scratch,
                                 Condition cc,
                                 Label* branch) {
   ASSERT(cc == equal || cc == not_equal);
   if (Serializer::enabled()) {
     // Can't do arithmetic on external references if it might get serialized.
     mov(scratch, Operand(object));
     // The mask isn't really an address.  We load it as an external reference in
     // case the size of the new space is different between the snapshot maker
     // and the running system.
     and_(Operand(scratch), Immediate(ExternalReference::new_space_mask()));
     cmp(Operand(scratch), Immediate(ExternalReference::new_space_start()));
     j(cc, branch);
   } else {
     int32_t new_space_start = reinterpret_cast<int32_t>(
         ExternalReference::new_space_start().address());
     lea(scratch, Operand(object, -new_space_start));
     and_(scratch, Heap::NewSpaceMask());
     j(cc, branch);
   }
 }
 
 
-// For page containing |object| mark region covering [object+offset] dirty.
+// Set the remembered set bit for [object+offset].
 // object is the object being stored into, value is the object being stored.
 // If offset is zero, then the scratch register contains the array index into
 // the elements array represented as a Smi.
 // All registers are clobbered by the operation.
 void MacroAssembler::RecordWrite(Register object, int offset,
                                  Register value, Register scratch) {
   // The compiled code assumes that record write doesn't change the
   // context register, so we check that none of the clobbered
   // registers are esi.
   ASSERT(!object.is(esi) && !value.is(esi) && !scratch.is(esi));
 
-  // First, check if a write barrier is even needed. The tests below
-  // catch stores of Smis and stores into young gen.
+  // First, check if a remembered set write is even needed. The tests below
+  // catch stores of Smis and stores into young gen (which does not have space
+  // for the remembered set bits).
   Label done;
 
   // Skip barrier if writing a smi.
   ASSERT_EQ(0, kSmiTag);
   test(value, Immediate(kSmiTagMask));
   j(zero, &done);
 
   InNewSpace(object, value, equal, &done);
 
   // The offset is relative to a tagged or untagged HeapObject pointer,
   // so either offset or offset + kHeapObjectTag must be a
   // multiple of kPointerSize.
   ASSERT(IsAligned(offset, kPointerSize) ||
          IsAligned(offset + kHeapObjectTag, kPointerSize));
 
-  Register dst = scratch;
-  if (offset != 0) {
-    lea(dst, Operand(object, offset));
+  // We use optimized write barrier code if the word being written to is not in
+  // a large object chunk or is in the first page of a large object chunk.
+  // We make sure that an offset is inside the right limits whether it is
+  // tagged or untagged.
+  if ((offset > 0) && (offset < Page::kMaxHeapObjectSize - kHeapObjectTag)) {
+    // Compute the bit offset in the remembered set, leave it in 'value'.
+    lea(value, Operand(object, offset));
+    and_(value, Page::kPageAlignmentMask);
+    shr(value, kPointerSizeLog2);
+
+    // Compute the page address from the heap object pointer, leave it in
+    // 'object'.
+    and_(object, ~Page::kPageAlignmentMask);
+
+    // NOTE: For now, we use the bit-test-and-set (bts) x86 instruction
+    // to limit code size. We should probably evaluate this decision by
+    // measuring the performance of an equivalent implementation using
+    // "simpler" instructions
+    bts(Operand(object, Page::kRSetOffset), value);
   } else {
-    // Array access: calculate the destination address in the same manner as
-    // KeyedStoreIC::GenerateGeneric.  Multiply a smi by 2 to get an offset
-    // into an array of words.
-    ASSERT_EQ(1, kSmiTagSize);
-    ASSERT_EQ(0, kSmiTag);
-    lea(dst, Operand(object, dst, times_half_pointer_size,
-                     FixedArray::kHeaderSize - kHeapObjectTag));
+    Register dst = scratch;
+    if (offset != 0) {
+      lea(dst, Operand(object, offset));
+    } else {
+      // array access: calculate the destination address in the same manner as
+      // KeyedStoreIC::GenerateGeneric.  Multiply a smi by 2 to get an offset
+      // into an array of words.
+      ASSERT_EQ(1, kSmiTagSize);
+      ASSERT_EQ(0, kSmiTag);
+      lea(dst, Operand(object, dst, times_half_pointer_size,
+                       FixedArray::kHeaderSize - kHeapObjectTag));
+    }
+    // If we are already generating a shared stub, not inlining the
+    // record write code isn't going to save us any memory.
+    if (generating_stub()) {
+      RecordWriteHelper(object, dst, value);
+    } else {
+      RecordWriteStub stub(object, dst, value);
+      CallStub(&stub);
+    }
   }
-  RecordWriteHelper(object, dst, value);
 
   bind(&done);
 
   // Clobber all input registers when running with the debug-code flag
   // turned on to provoke errors.
   if (FLAG_debug_code) {
     mov(object, Immediate(BitCast<int32_t>(kZapValue)));
     mov(value, Immediate(BitCast<int32_t>(kZapValue)));
     mov(scratch, Immediate(BitCast<int32_t>(kZapValue)));
   }
@@ skipping 1163 matching lines @@
 }
 
 
 void MacroAssembler::InvokeFunction(Register fun,
                                     const ParameterCount& actual,
                                     InvokeFlag flag) {
   ASSERT(fun.is(edi));
   mov(edx, FieldOperand(edi, JSFunction::kSharedFunctionInfoOffset));
   mov(esi, FieldOperand(edi, JSFunction::kContextOffset));
   mov(ebx, FieldOperand(edx, SharedFunctionInfo::kFormalParameterCountOffset));
-  SmiUntag(ebx);
   mov(edx, FieldOperand(edx, SharedFunctionInfo::kCodeOffset));
   lea(edx, FieldOperand(edx, Code::kHeaderSize));
 
   ParameterCount expected(ebx);
   InvokeCode(Operand(edx), expected, actual, flag);
 }
 
 
 void MacroAssembler::InvokeFunction(JSFunction* function,
                                     const ParameterCount& actual,
@@ skipping 306 matching lines @@
 
   // Check that the code was patched as expected.
   ASSERT(masm_.pc_ == address_ + size_);
   ASSERT(masm_.reloc_info_writer.pos() == address_ + size_ + Assembler::kGap);
 }
 
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_IA32