Use nearlabel AFAP in lithium codegen

src/ia32/lithium-codegen-ia32.cc

 // Copyright 2012 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 2029 matching lines @@
   if (index->value() == 0) {
     __ mov(result, FieldOperand(object, JSDate::kValueOffset));
   } else {
     if (index->value() < JSDate::kFirstUncachedField) {
       ExternalReference stamp = ExternalReference::date_cache_stamp(isolate());
       __ mov(scratch, Operand::StaticVariable(stamp));
       __ cmp(scratch, FieldOperand(object, JSDate::kCacheStampOffset));
       __ j(not_equal, &runtime, Label::kNear);
       __ mov(result, FieldOperand(object, JSDate::kValueOffset +
                                           kPointerSize * index->value()));
-      __ jmp(&done);
+      __ jmp(&done, Label::kNear);
     }
     __ bind(&runtime);
     __ PrepareCallCFunction(2, scratch);
     __ mov(Operand(esp, 0), object);
     __ mov(Operand(esp, 1 * kPointerSize), Immediate(index));
     __ CallCFunction(ExternalReference::get_date_field_function(isolate()), 2);
     __ bind(&done);
   }
 }
 
@@ skipping 521 matching lines @@
     __ ucomisd(input_reg, input_reg);
     EmitFalseBranch(instr, parity_odd);
   } else {
     // Put the value to the top of stack
     X87Register src = ToX87Register(instr->object());
     X87LoadForUsage(src);
     __ fld(0);
     __ fld(0);
     __ FCmp();
     Label ok;
-    __ j(parity_even, &ok);
+    __ j(parity_even, &ok, Label::kNear);
     __ fstp(0);
     EmitFalseBranch(instr, no_condition);
     __ bind(&ok);
   }
 
 
   __ sub(esp, Immediate(kDoubleSize));
   if (use_sse2) {
     CpuFeatureScope scope(masm(), SSE2);
     XMMRegister input_reg = ToDoubleRegister(instr->object());
@@ skipping 303 matching lines @@
   };
 
   DeferredInstanceOfKnownGlobal* deferred;
   deferred = new(zone()) DeferredInstanceOfKnownGlobal(this, instr, x87_stack_);
 
   Label done, false_result;
   Register object = ToRegister(instr->value());
   Register temp = ToRegister(instr->temp());
 
   // A Smi is not an instance of anything.
-  __ JumpIfSmi(object, &false_result);
+  __ JumpIfSmi(object, &false_result, Label::kNear);
 
   // This is the inlined call site instanceof cache. The two occurences of the
   // hole value will be patched to the last map/result pair generated by the
   // instanceof stub.
   Label cache_miss;
   Register map = ToRegister(instr->temp());
   __ mov(map, FieldOperand(object, HeapObject::kMapOffset));
   __ bind(deferred->map_check());  // Label for calculating code patching.
   Handle<Cell> cache_cell = factory()->NewCell(factory()->the_hole_value());
   __ cmp(map, Operand::ForCell(cache_cell));  // Patched to cached map.
   __ j(not_equal, &cache_miss, Label::kNear);
   __ mov(eax, factory()->the_hole_value());  // Patched to either true or false.
-  __ jmp(&done);
+  __ jmp(&done, Label::kNear);
 
   // The inlined call site cache did not match. Check for null and string
   // before calling the deferred code.
   __ bind(&cache_miss);
   // Null is not an instance of anything.
   __ cmp(object, factory()->null_value());
-  __ j(equal, &false_result);
+  __ j(equal, &false_result, Label::kNear);
 
   // String values are not instances of anything.
   Condition is_string = masm_->IsObjectStringType(object, temp, temp);
-  __ j(is_string, &false_result);
+  __ j(is_string, &false_result, Label::kNear);
 
   // Go to the deferred code.
   __ jmp(deferred->entry());
 
   __ bind(&false_result);
   __ mov(ToRegister(instr->result()), factory()->false_value());
 
   // Here result has either true or false. Deferred code also produces true or
   // false object.
   __ bind(deferred->exit());
@@ skipping 124 matching lines @@
   }
   int no_frame_start = -1;
   if (NeedsEagerFrame()) {
     __ mov(esp, ebp);
     __ pop(ebp);
     no_frame_start = masm_->pc_offset();
   }
   if (dynamic_frame_alignment_) {
     Label no_padding;
     __ cmp(edx, Immediate(kNoAlignmentPadding));
-    __ j(equal, &no_padding);
+    __ j(equal, &no_padding, Label::kNear);
 
     EmitReturn(instr, true);
     __ bind(&no_padding);
   }
 
   EmitReturn(instr, false);
   if (no_frame_start != -1) {
     info()->AddNoFrameRange(no_frame_start, masm_->pc_offset());
   }
 }
@@ skipping 504 matching lines @@
   // object as a receiver to normal functions. Values have to be
   // passed unchanged to builtins and strict-mode functions.
   Label global_object, receiver_ok;
 
   // Do not transform the receiver to object for strict mode
   // functions.
   __ mov(scratch,
          FieldOperand(function, JSFunction::kSharedFunctionInfoOffset));
   __ test_b(FieldOperand(scratch, SharedFunctionInfo::kStrictModeByteOffset),
             1 << SharedFunctionInfo::kStrictModeBitWithinByte);
-  __ j(not_equal, &receiver_ok);  // A near jump is not sufficient here!

[Jakob Kummerow, 2013/10/28 10:37:19]

Why do you think this comment is not true anymore?

[Weiliang, 2013/10/28 14:31:49]

Yes, thanks for pointing it out. near jump is not sufficient here due to
deopt-n-times flags. 

Seems current test does not find out this bug

+  __ j(not_equal, &receiver_ok, Label::kNear);
 
   // Do not transform the receiver to object for builtins.
   __ test_b(FieldOperand(scratch, SharedFunctionInfo::kNativeByteOffset),
             1 << SharedFunctionInfo::kNativeBitWithinByte);
-  __ j(not_equal, &receiver_ok);
+  __ j(not_equal, &receiver_ok, Label::kNear);
 
   // Normal function. Replace undefined or null with global receiver.
   __ cmp(receiver, factory()->null_value());
   __ j(equal, &global_object, Label::kNear);
   __ cmp(receiver, factory()->undefined_value());
   __ j(equal, &global_object, Label::kNear);
 
   // The receiver should be a JS object.
   __ test(receiver, Immediate(kSmiTagMask));
   DeoptimizeIf(equal, instr->environment());
@@ skipping 188 matching lines @@
 
   // Preserve the value of all registers.
   PushSafepointRegistersScope scope(this);
 
   __ mov(tmp, FieldOperand(input_reg, HeapNumber::kExponentOffset));
   // Check the sign of the argument. If the argument is positive, just
   // return it. We do not need to patch the stack since |input| and
   // |result| are the same register and |input| will be restored
   // unchanged by popping safepoint registers.
   __ test(tmp, Immediate(HeapNumber::kSignMask));
-  __ j(zero, &done);
+  __ j(zero, &done, Label::kNear);
 
   __ AllocateHeapNumber(tmp, tmp2, no_reg, &slow);
   __ jmp(&allocated, Label::kNear);
 
   // Slow case: Call the runtime system to do the number allocation.
   __ bind(&slow);
   CallRuntimeFromDeferred(Runtime::kAllocateHeapNumber, 0,
                           instr, instr->context());
   // Set the pointer to the new heap number in tmp.
   if (!tmp.is(eax)) __ mov(tmp, eax);
@@ skipping 135 matching lines @@
   XMMRegister input_reg = ToDoubleRegister(instr->value());
   XMMRegister xmm_scratch = double_scratch0();
   XMMRegister input_temp = ToDoubleRegister(instr->temp());
   ExternalReference one_half = ExternalReference::address_of_one_half();
   ExternalReference minus_one_half =
       ExternalReference::address_of_minus_one_half();
 
   Label done, round_to_zero, below_one_half, do_not_compensate;
   __ movsd(xmm_scratch, Operand::StaticVariable(one_half));
   __ ucomisd(xmm_scratch, input_reg);
-  __ j(above, &below_one_half);
+  __ j(above, &below_one_half, Label::kNear);
 
   // CVTTSD2SI rounds towards zero, since 0.5 <= x, we use floor(0.5 + x).
   __ addsd(xmm_scratch, input_reg);
   __ cvttsd2si(output_reg, Operand(xmm_scratch));
   // Overflow is signalled with minint.
   __ cmp(output_reg, 0x80000000u);
   __ RecordComment("D2I conversion overflow");
   DeoptimizeIf(equal, instr->environment());
-  __ jmp(&done);
+  __ jmp(&done, Label::kNear);
 
   __ bind(&below_one_half);
   __ movsd(xmm_scratch, Operand::StaticVariable(minus_one_half));
   __ ucomisd(xmm_scratch, input_reg);
-  __ j(below_equal, &round_to_zero);
+  __ j(below_equal, &round_to_zero, Label::kNear);
 
   // CVTTSD2SI rounds towards zero, we use ceil(x - (-0.5)) and then
   // compare and compensate.
   __ movsd(input_temp, input_reg);  // Do not alter input_reg.
   __ subsd(input_temp, xmm_scratch);
   __ cvttsd2si(output_reg, Operand(input_temp));
   // Catch minint due to overflow, and to prevent overflow when compensating.
   __ cmp(output_reg, 0x80000000u);
   __ RecordComment("D2I conversion overflow");
   DeoptimizeIf(equal, instr->environment());
 
   __ Cvtsi2sd(xmm_scratch, output_reg);
   __ ucomisd(xmm_scratch, input_temp);
-  __ j(equal, &done);
+  __ j(equal, &done, Label::kNear);
   __ sub(output_reg, Immediate(1));
   // No overflow because we already ruled out minint.
-  __ jmp(&done);
+  __ jmp(&done, Label::kNear);
 
   __ bind(&round_to_zero);
   // We return 0 for the input range [+0, 0.5[, or [-0.5, 0.5[ if
   // we can ignore the difference between a result of -0 and +0.
   if (instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero)) {
     // If the sign is positive, we return +0.
     __ movmskpd(output_reg, input_reg);
     __ test(output_reg, Immediate(1));
     __ RecordComment("Minus zero");
     DeoptimizeIf(not_zero, instr->environment());
@@ skipping 334 matching lines @@
     ArrayNoArgumentConstructorStub stub(kind, context_mode, override_mode);
     CallCode(stub.GetCode(isolate()), RelocInfo::CONSTRUCT_CALL, instr);
   } else if (instr->arity() == 1) {
     Label done;
     if (IsFastPackedElementsKind(kind)) {
       Label packed_case;
       // We might need a change here
       // look at the first argument
       __ mov(ecx, Operand(esp, 0));
       __ test(ecx, ecx);
-      __ j(zero, &packed_case);
+      __ j(zero, &packed_case, Label::kNear);
 
       ElementsKind holey_kind = GetHoleyElementsKind(kind);
       ArraySingleArgumentConstructorStub stub(holey_kind, context_mode,
                                               override_mode);
       CallCode(stub.GetCode(isolate()), RelocInfo::CONSTRUCT_CALL, instr);
-      __ jmp(&done);
+      __ jmp(&done, Label::kNear);
       __ bind(&packed_case);
     }
 
     ArraySingleArgumentConstructorStub stub(kind, context_mode, override_mode);
     CallCode(stub.GetCode(isolate()), RelocInfo::CONSTRUCT_CALL, instr);
     __ bind(&done);
   } else {
     ArrayNArgumentsConstructorStub stub(kind, context_mode, override_mode);
     CallCode(stub.GetCode(isolate()), RelocInfo::CONSTRUCT_CALL, instr);
   }
@@ skipping 276 matching lines @@
       instr->additional_index());
 
   if (CpuFeatures::IsSafeForSnapshot(SSE2)) {
     CpuFeatureScope scope(masm(), SSE2);
     XMMRegister value = ToDoubleRegister(instr->value());
 
     if (instr->NeedsCanonicalization()) {
       Label have_value;
 
       __ ucomisd(value, value);
-      __ j(parity_odd, &have_value);  // NaN.
+      __ j(parity_odd, &have_value, Label::kNear);  // NaN.
 
       __ movsd(value, Operand::StaticVariable(canonical_nan_reference));
       __ bind(&have_value);
     }
 
     __ movsd(double_store_operand, value);
   } else {
     // Can't use SSE2 in the serializer
     if (instr->hydrogen()->IsConstantHoleStore()) {
       // This means we should store the (double) hole. No floating point
@@ skipping 15 matching lines @@
     } else {
       Label no_special_nan_handling;
       X87Register value = ToX87Register(instr->value());
       X87Fxch(value);
 
       if (instr->NeedsCanonicalization()) {
         __ fld(0);
         __ fld(0);
         __ FCmp();
 
-        __ j(parity_odd, &no_special_nan_handling);
+        __ j(parity_odd, &no_special_nan_handling, Label::kNear);
         __ sub(esp, Immediate(kDoubleSize));
         __ fst_d(MemOperand(esp, 0));
         __ cmp(MemOperand(esp, sizeof(kHoleNanLower32)),
                Immediate(kHoleNanUpper32));
         __ add(esp, Immediate(kDoubleSize));
         Label canonicalize;
-        __ j(not_equal, &canonicalize);
-        __ jmp(&no_special_nan_handling);
+        __ j(not_equal, &canonicalize, Label::kNear);
+        __ jmp(&no_special_nan_handling, Label::kNear);
         __ bind(&canonicalize);
         __ fstp(0);
         __ fld_d(Operand::StaticVariable(canonical_nan_reference));
       }
 
       __ bind(&no_special_nan_handling);
       __ fst_d(double_store_operand);
     }
   }
 }
@@ skipping 948 matching lines @@
   DeferredCheckMaps* deferred = NULL;
   if (instr->hydrogen()->has_migration_target()) {
     deferred = new(zone()) DeferredCheckMaps(this, instr, reg, x87_stack_);
     __ bind(deferred->check_maps());
   }
 
   UniqueSet<Map> map_set = instr->hydrogen()->map_set();
   Label success;
   for (int i = 0; i < map_set.size() - 1; i++) {
     Handle<Map> map = map_set.at(i).handle();
-    __ CompareMap(reg, map, &success);
-    __ j(equal, &success);
+    __ CompareMap(reg, map);
+    __ j(equal, &success, Label::kNear);
   }
 
   Handle<Map> map = map_set.at(map_set.size() - 1).handle();
-  __ CompareMap(reg, map, &success);
+  __ CompareMap(reg, map);
   if (instr->hydrogen()->has_migration_target()) {
     __ j(not_equal, deferred->entry());
   } else {
     DeoptimizeIf(not_equal, instr->environment());
   }
 
   __ bind(&success);
 }
 
 
@@ skipping 57 matching lines @@
   Register scratch2 = ToRegister(instr->scratch2());
   Register scratch3 = ToRegister(instr->scratch3());
   Label is_smi, done, heap_number, valid_exponent,
       largest_value, zero_result, maybe_nan_or_infinity;
 
   __ JumpIfSmi(input_reg, &is_smi);
 
   // Check for heap number
   __ cmp(FieldOperand(input_reg, HeapObject::kMapOffset),
          factory()->heap_number_map());
-  __ j(equal, &heap_number, Label::kFar);
+  __ j(equal, &heap_number, Label::kNear);
 
   // Check for undefined. Undefined is converted to zero for clamping
   // conversions.
   __ cmp(input_reg, factory()->undefined_value());
   DeoptimizeIf(not_equal, instr->environment());
-  __ jmp(&zero_result);
+  __ jmp(&zero_result, Label::kNear);
 
   // Heap number
   __ bind(&heap_number);
 
   // Surprisingly, all of the hand-crafted bit-manipulations below are much
   // faster than the x86 FPU built-in instruction, especially since "banker's
   // rounding" would be additionally very expensive
 
   // Get exponent word.
   __ mov(scratch, FieldOperand(input_reg, HeapNumber::kExponentOffset));
   __ mov(scratch3, FieldOperand(input_reg, HeapNumber::kMantissaOffset));
 
   // Test for negative values --> clamp to zero
   __ test(scratch, scratch);
-  __ j(negative, &zero_result);
+  __ j(negative, &zero_result, Label::kNear);
 
   // Get exponent alone in scratch2.
   __ mov(scratch2, scratch);
   __ and_(scratch2, HeapNumber::kExponentMask);
   __ shr(scratch2, HeapNumber::kExponentShift);
-  __ j(zero, &zero_result);
+  __ j(zero, &zero_result, Label::kNear);
   __ sub(scratch2, Immediate(HeapNumber::kExponentBias - 1));
-  __ j(negative, &zero_result);
+  __ j(negative, &zero_result, Label::kNear);
 
   const uint32_t non_int8_exponent = 7;
   __ cmp(scratch2, Immediate(non_int8_exponent + 1));
   // If the exponent is too big, check for special values.
   __ j(greater, &maybe_nan_or_infinity, Label::kNear);
 
   __ bind(&valid_exponent);
   // Exponent word in scratch, exponent in scratch2. We know that 0 <= exponent
   // < 7. The shift bias is the number of bits to shift the mantissa such that
   // with an exponent of 7 such the that top-most one is in bit 30, allowing
@@ skipping 10 matching lines @@
   __ shl_cl(scratch);
   // Use "banker's rounding" to spec: If fractional part of number is 0.5, then
   // use the bit in the "ones" place and add it to the "halves" place, which has
   // the effect of rounding to even.
   __ mov(scratch2, scratch);
   const uint32_t one_half_bit_shift = 30 - sizeof(uint8_t) * 8;
   const uint32_t one_bit_shift = one_half_bit_shift + 1;
   __ and_(scratch2, Immediate((1 << one_bit_shift) - 1));
   __ cmp(scratch2, Immediate(1 << one_half_bit_shift));
   Label no_round;
-  __ j(less, &no_round);
+  __ j(less, &no_round, Label::kNear);
   Label round_up;
   __ mov(scratch2, Immediate(1 << one_half_bit_shift));
-  __ j(greater, &round_up);
+  __ j(greater, &round_up, Label::kNear);
   __ test(scratch3, scratch3);
-  __ j(not_zero, &round_up);
+  __ j(not_zero, &round_up, Label::kNear);
   __ mov(scratch2, scratch);
   __ and_(scratch2, Immediate(1 << one_bit_shift));
   __ shr(scratch2, 1);
   __ bind(&round_up);
   __ add(scratch, scratch2);
-  __ j(overflow, &largest_value);
+  __ j(overflow, &largest_value, Label::kNear);
   __ bind(&no_round);
   __ shr(scratch, 23);
   __ mov(result_reg, scratch);
   __ jmp(&done, Label::kNear);
 
   __ bind(&maybe_nan_or_infinity);
   // Check for NaN/Infinity, all other values map to 255
   __ cmp(scratch2, Immediate(HeapNumber::kInfinityOrNanExponent + 1));
   __ j(not_equal, &largest_value, Label::kNear);
 
   // Check for NaN, which differs from Infinity in that at least one mantissa
   // bit is set.
   __ and_(scratch, HeapNumber::kMantissaMask);
   __ or_(scratch, FieldOperand(input_reg, HeapNumber::kMantissaOffset));
-  __ j(not_zero, &zero_result);  // M!=0 --> NaN
+  __ j(not_zero, &zero_result, Label::kNear);  // M!=0 --> NaN
   // Infinity -> Fall through to map to 255.
 
   __ bind(&largest_value);
   __ mov(result_reg, Immediate(255));
   __ jmp(&done, Label::kNear);
 
   __ bind(&zero_result);
   __ xor_(result_reg, result_reg);
-  __ jmp(&done);
+  __ jmp(&done, Label::kNear);
 
   // smi
   __ bind(&is_smi);
   if (!input_reg.is(result_reg)) {
     __ mov(result_reg, input_reg);
   }
   __ SmiUntag(result_reg);
   __ ClampUint8(result_reg);
   __ bind(&done);
 }
@@ skipping 127 matching lines @@
   __ push(Immediate(Smi::FromInt(instr->hydrogen()->literal_index())));
   __ push(Immediate(instr->hydrogen()->pattern()));
   __ push(Immediate(instr->hydrogen()->flags()));
   CallRuntime(Runtime::kMaterializeRegExpLiteral, 4, instr);
   __ mov(ebx, eax);
 
   __ bind(&materialized);
   int size = JSRegExp::kSize + JSRegExp::kInObjectFieldCount * kPointerSize;
   Label allocated, runtime_allocate;
   __ Allocate(size, eax, ecx, edx, &runtime_allocate, TAG_OBJECT);
-  __ jmp(&allocated);
+  __ jmp(&allocated, Label::kNear);
 
   __ bind(&runtime_allocate);
   __ push(ebx);
   __ push(Immediate(Smi::FromInt(size)));
   CallRuntime(Runtime::kAllocateInNewSpace, 1, instr);
   __ pop(ebx);
 
   __ bind(&allocated);
   // Copy the content into the newly allocated memory.
   // (Unroll copy loop once for better throughput).
@@ skipping 305 matching lines @@
   __ bind(&use_cache);
 }
 
 
 void LCodeGen::DoForInCacheArray(LForInCacheArray* instr) {
   Register map = ToRegister(instr->map());
   Register result = ToRegister(instr->result());
   Label load_cache, done;
   __ EnumLength(result, map);
   __ cmp(result, Immediate(Smi::FromInt(0)));
-  __ j(not_equal, &load_cache);
+  __ j(not_equal, &load_cache, Label::kNear);
   __ mov(result, isolate()->factory()->empty_fixed_array());
-  __ jmp(&done);
+  __ jmp(&done, Label::kNear);
 
   __ bind(&load_cache);
   __ LoadInstanceDescriptors(map, result);
   __ mov(result,
          FieldOperand(result, DescriptorArray::kEnumCacheOffset));
   __ mov(result,
          FieldOperand(result, FixedArray::SizeFor(instr->idx())));
   __ bind(&done);
   __ test(result, result);
   DeoptimizeIf(equal, instr->environment());
 }
 
 
 void LCodeGen::DoCheckMapValue(LCheckMapValue* instr) {
   Register object = ToRegister(instr->value());
   __ cmp(ToRegister(instr->map()),
          FieldOperand(object, HeapObject::kMapOffset));
   DeoptimizeIf(not_equal, instr->environment());
 }
 
 
 void LCodeGen::DoLoadFieldByIndex(LLoadFieldByIndex* instr) {
   Register object = ToRegister(instr->object());
   Register index = ToRegister(instr->index());
 
   Label out_of_object, done;
   __ cmp(index, Immediate(0));
-  __ j(less, &out_of_object);
+  __ j(less, &out_of_object, Label::kNear);
   __ mov(object, FieldOperand(object,
                               index,
                               times_half_pointer_size,
                               JSObject::kHeaderSize));
   __ jmp(&done, Label::kNear);
 
   __ bind(&out_of_object);
   __ mov(object, FieldOperand(object, JSObject::kPropertiesOffset));
   __ neg(index);
   // Index is now equal to out of object property index plus 1.
   __ mov(object, FieldOperand(object,
                               index,
                               times_half_pointer_size,
                               FixedArray::kHeaderSize - kPointerSize));
   __ bind(&done);
 }
 
 
 #undef __
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_IA32