DoNumberTagD performance improvement

Allocate heap entry untagged and tag at end to avoid having to subtract off
the tag offset before storing the value.

BUG=

Review URL: https://codereview.chromium.org/11028115
Patch from Anthony Berent <aberent@chromium.org>.

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12711 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/arm/lithium-codegen-arm.cc

@@ -4446,7 +4446,7 @@ void LCodeGen::DoDeferredNumberTagI(LInstruction* instr,
 
   if (FLAG_inline_new) {
     __ LoadRoot(r6, Heap::kHeapNumberMapRootIndex);
-    __ AllocateHeapNumber(r5, r3, r4, r6, &slow);
+    __ AllocateHeapNumber(r5, r3, r4, r6, &slow, DONT_TAG_RESULT);
     __ Move(dst, r5);
     __ b(&done);
   }
@@ -4461,12 +4461,13 @@ void LCodeGen::DoDeferredNumberTagI(LInstruction* instr,
   __ StoreToSafepointRegisterSlot(ip, dst);
   CallRuntimeFromDeferred(Runtime::kAllocateHeapNumber, 0, instr);
   __ Move(dst, r0);
+  __ sub(dst, dst, Operand(kHeapObjectTag));
 
   // Done. Put the value in dbl_scratch into the value of the allocated heap
   // number.
   __ bind(&done);
-  __ sub(ip, dst, Operand(kHeapObjectTag));
-  __ vstr(dbl_scratch, ip, HeapNumber::kValueOffset);
+  __ vstr(dbl_scratch, dst, HeapNumber::kValueOffset);
+  __ add(dst, dst, Operand(kHeapObjectTag));
   __ StoreToSafepointRegisterSlot(dst, dst);
 }
 
@@ -4491,13 +4492,16 @@ void LCodeGen::DoNumberTagD(LNumberTagD* instr) {
   DeferredNumberTagD* deferred = new(zone()) DeferredNumberTagD(this, instr);
   if (FLAG_inline_new) {
     __ LoadRoot(scratch, Heap::kHeapNumberMapRootIndex);
-    __ AllocateHeapNumber(reg, temp1, temp2, scratch, deferred->entry());
+    // We want the untagged address first for performance
+    __ AllocateHeapNumber(reg, temp1, temp2, scratch, deferred->entry(),
+                          DONT_TAG_RESULT);
   } else {
     __ jmp(deferred->entry());
   }
   __ bind(deferred->exit());
-  __ sub(ip, reg, Operand(kHeapObjectTag));
-  __ vstr(input_reg, ip, HeapNumber::kValueOffset);
+  __ vstr(input_reg, reg, HeapNumber::kValueOffset);
+  // Now that we have finished with the object's real address tag it
+  __ add(reg, reg, Operand(kHeapObjectTag));
 }
 
 
@@ -4510,6 +4514,7 @@ void LCodeGen::DoDeferredNumberTagD(LNumberTagD* instr) {
 
   PushSafepointRegistersScope scope(this, Safepoint::kWithRegisters);
   CallRuntimeFromDeferred(Runtime::kAllocateHeapNumber, 0, instr);
+  __ sub(r0, r0, Operand(kHeapObjectTag));
   __ StoreToSafepointRegisterSlot(r0, reg);
 }
 

src/arm/macro-assembler-arm.cc

@@ -3124,7 +3124,8 @@ void MacroAssembler::AllocateHeapNumber(Register result,
                                         Register scratch1,
                                         Register scratch2,
                                         Register heap_number_map,
-                                        Label* gc_required) {
+                                        Label* gc_required,
+                                        TaggingMode tagging_mode) {
   // Allocate an object in the heap for the heap number and tag it as a heap
   // object.
   AllocateInNewSpace(HeapNumber::kSize,
@@ -3132,11 +3133,16 @@ void MacroAssembler::AllocateHeapNumber(Register result,
                      scratch1,
                      scratch2,
                      gc_required,
-                     TAG_OBJECT);
+                     tagging_mode == TAG_RESULT ? TAG_OBJECT :
+                                                  NO_ALLOCATION_FLAGS);
 
   // Store heap number map in the allocated object.
   AssertRegisterIsRoot(heap_number_map, Heap::kHeapNumberMapRootIndex);
-  str(heap_number_map, FieldMemOperand(result, HeapObject::kMapOffset));
+  if (tagging_mode == TAG_RESULT) {
+    str(heap_number_map, FieldMemOperand(result, HeapObject::kMapOffset));
+  } else {
+    str(heap_number_map, MemOperand(result, HeapObject::kMapOffset));
+  }
 }
 
 

src/arm/macro-assembler-arm.h

@@ -68,6 +68,13 @@ enum AllocationFlags {
   SIZE_IN_WORDS = 1 << 2
 };
 
+// Flags used for AllocateHeapNumber
+enum TaggingMode {
+  // Tag the result.
+  TAG_RESULT,
+  // Don't tag
+  DONT_TAG_RESULT
+};
 
 // Flags used for the ObjectToDoubleVFPRegister function.
 enum ObjectToDoubleFlags {
@@ -731,7 +738,8 @@ class MacroAssembler: public Assembler {
                           Register scratch1,
                           Register scratch2,
                           Register heap_number_map,
-                          Label* gc_required);
+                          Label* gc_required,
+                          TaggingMode tagging_mode = TAG_RESULT);
   void AllocateHeapNumberWithValue(Register result,
                                    DwVfpRegister value,
                                    Register scratch1,

src/arm/stub-cache-arm.cc

@@ -3787,22 +3787,28 @@ void KeyedLoadStubCompiler::GenerateLoadExternalArray(
     __ Ret();
 
     __ bind(&box_int);
-    // Allocate a HeapNumber for the result and perform int-to-double
-    // conversion.  Don't touch r0 or r1 as they are needed if allocation
-    // fails.
-    __ LoadRoot(r6, Heap::kHeapNumberMapRootIndex);
-    __ AllocateHeapNumber(r5, r3, r4, r6, &slow);
-    // Now we can use r0 for the result as key is not needed any more.
-    __ mov(r0, r5);
-
     if (CpuFeatures::IsSupported(VFP2)) {
       CpuFeatures::Scope scope(VFP2);
+      // Allocate a HeapNumber for the result and perform int-to-double
+      // conversion.  Don't touch r0 or r1 as they are needed if allocation
+      // fails.
+      __ LoadRoot(r6, Heap::kHeapNumberMapRootIndex);
+
+      __ AllocateHeapNumber(r5, r3, r4, r6, &slow, DONT_TAG_RESULT);
+      // Now we can use r0 for the result as key is not needed any more.
+      __ add(r0, r5, Operand(kHeapObjectTag));
       __ vmov(s0, value);
       __ vcvt_f64_s32(d0, s0);
-      __ sub(r3, r0, Operand(kHeapObjectTag));
-      __ vstr(d0, r3, HeapNumber::kValueOffset);
+      __ vstr(d0, r5, HeapNumber::kValueOffset);
       __ Ret();
     } else {
+      // Allocate a HeapNumber for the result and perform int-to-double
+      // conversion.  Don't touch r0 or r1 as they are needed if allocation
+      // fails.
+      __ LoadRoot(r6, Heap::kHeapNumberMapRootIndex);
+      __ AllocateHeapNumber(r5, r3, r4, r6, &slow, TAG_RESULT);
+      // Now we can use r0 for the result as key is not needed any more.
+      __ mov(r0, r5);
       Register dst1 = r1;
       Register dst2 = r3;
       FloatingPointHelper::Destination dest =
@@ -3838,13 +3844,12 @@ void KeyedLoadStubCompiler::GenerateLoadExternalArray(
       // conversion. Don't use r0 and r1 as AllocateHeapNumber clobbers all
       // registers - also when jumping due to exhausted young space.
       __ LoadRoot(r6, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(r2, r3, r4, r6, &slow);
+      __ AllocateHeapNumber(r2, r3, r4, r6, &slow, DONT_TAG_RESULT);
 
       __ vcvt_f64_u32(d0, s0);
-      __ sub(r1, r2, Operand(kHeapObjectTag));
-      __ vstr(d0, r1, HeapNumber::kValueOffset);
+      __ vstr(d0, r2, HeapNumber::kValueOffset);
 
-      __ mov(r0, r2);
+      __ add(r0, r2, Operand(kHeapObjectTag));
       __ Ret();
     } else {
       // Check whether unsigned integer fits into smi.
@@ -3876,7 +3881,7 @@ void KeyedLoadStubCompiler::GenerateLoadExternalArray(
       // clobbers all registers - also when jumping due to exhausted young
       // space.
       __ LoadRoot(r6, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(r4, r5, r7, r6, &slow);
+      __ AllocateHeapNumber(r4, r5, r7, r6, &slow, TAG_RESULT);
 
       __ str(hiword, FieldMemOperand(r4, HeapNumber::kExponentOffset));
       __ str(loword, FieldMemOperand(r4, HeapNumber::kMantissaOffset));
@@ -3893,19 +3898,18 @@ void KeyedLoadStubCompiler::GenerateLoadExternalArray(
       // AllocateHeapNumber clobbers all registers - also when jumping due to
       // exhausted young space.
       __ LoadRoot(r6, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(r2, r3, r4, r6, &slow);
+      __ AllocateHeapNumber(r2, r3, r4, r6, &slow, DONT_TAG_RESULT);
       __ vcvt_f64_f32(d0, s0);
-      __ sub(r1, r2, Operand(kHeapObjectTag));
-      __ vstr(d0, r1, HeapNumber::kValueOffset);
+      __ vstr(d0, r2, HeapNumber::kValueOffset);
 
-      __ mov(r0, r2);
+      __ add(r0, r2, Operand(kHeapObjectTag));
       __ Ret();
     } else {
       // Allocate a HeapNumber for the result. Don't use r0 and r1 as
       // AllocateHeapNumber clobbers all registers - also when jumping due to
       // exhausted young space.
       __ LoadRoot(r6, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(r3, r4, r5, r6, &slow);
+      __ AllocateHeapNumber(r3, r4, r5, r6, &slow, TAG_RESULT);
       // VFP is not available, do manual single to double conversion.
 
       // r2: floating point value (binary32)
@@ -3961,18 +3965,17 @@ void KeyedLoadStubCompiler::GenerateLoadExternalArray(
       // AllocateHeapNumber clobbers all registers - also when jumping due to
       // exhausted young space.
       __ LoadRoot(r6, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(r2, r3, r4, r6, &slow);
-      __ sub(r1, r2, Operand(kHeapObjectTag));
-      __ vstr(d0, r1, HeapNumber::kValueOffset);
+      __ AllocateHeapNumber(r2, r3, r4, r6, &slow, DONT_TAG_RESULT);
+      __ vstr(d0, r2, HeapNumber::kValueOffset);
 
-      __ mov(r0, r2);
+      __ add(r0, r2, Operand(kHeapObjectTag));
       __ Ret();
     } else {
       // Allocate a HeapNumber for the result. Don't use r0 and r1 as
       // AllocateHeapNumber clobbers all registers - also when jumping due to
       // exhausted young space.
       __ LoadRoot(r7, Heap::kHeapNumberMapRootIndex);
-      __ AllocateHeapNumber(r4, r5, r6, r7, &slow);
+      __ AllocateHeapNumber(r4, r5, r6, r7, &slow, TAG_RESULT);
 
       __ str(r2, FieldMemOperand(r4, HeapNumber::kMantissaOffset));
       __ str(r3, FieldMemOperand(r4, HeapNumber::kExponentOffset));
@@ -4439,7 +4442,7 @@ void KeyedLoadStubCompiler::GenerateLoadFastDoubleElement(
   // Non-NaN. Allocate a new heap number and copy the double value into it.
   __ LoadRoot(heap_number_map, Heap::kHeapNumberMapRootIndex);
   __ AllocateHeapNumber(heap_number_reg, scratch2, scratch3,
-                        heap_number_map, &slow_allocate_heapnumber);
+                        heap_number_map, &slow_allocate_heapnumber, TAG_RESULT);
 
   // Don't need to reload the upper 32 bits of the double, it's already in
   // scratch.