Some assembler-level optimizations on ARM.

src/arm/codegen-arm.cc

 // Copyright 2011 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 86 matching lines @@
   __ ldr(r4, FieldMemOperand(r2, JSObject::kElementsOffset));
   __ ldr(r5, FieldMemOperand(r4, FixedArray::kLengthOffset));
   // r4: source FixedArray
   // r5: number of elements (smi-tagged)
 
   // Allocate new FixedDoubleArray.
   __ mov(lr, Operand(FixedDoubleArray::kHeaderSize));
   __ add(lr, lr, Operand(r5, LSL, 2));
   __ AllocateInNewSpace(lr, r6, r7, r9, &gc_required, NO_ALLOCATION_FLAGS);
   // r6: destination FixedDoubleArray, not tagged as heap object
+  // Set destination FixedDoubleArray's length and map.
   __ LoadRoot(r9, Heap::kFixedDoubleArrayMapRootIndex);
+  __ str(r5, MemOperand(r6, FixedDoubleArray::kLengthOffset));
   __ str(r9, MemOperand(r6, HeapObject::kMapOffset));
-  // Set destination FixedDoubleArray's length.
-  __ str(r5, MemOperand(r6, FixedDoubleArray::kLengthOffset));
   // Update receiver's map.
 
   __ str(r3, FieldMemOperand(r2, HeapObject::kMapOffset));
   __ RecordWriteField(r2,
                       HeapObject::kMapOffset,
                       r3,
                       r9,
                       kLRHasBeenSaved,
                       kDontSaveFPRegs,
                       EMIT_REMEMBERED_SET,
@@ skipping 27 matching lines @@
 
   // Call into runtime if GC is required.
   __ bind(&gc_required);
   __ pop(lr);
   __ b(fail);
 
   // Convert and copy elements.
   __ bind(&loop);
   __ ldr(r9, MemOperand(r3, 4, PostIndex));
   // r9: current element
-  __ JumpIfNotSmi(r9, &convert_hole);
+  STATIC_ASSERT(kSmiTagSize == 1 && kSmiTag == 0);

[ulan, 2012/01/27 13:19:11]

Can we introduce UntagAndJumpIfNotSmi for these tree lines?

+  __ mov(r9, Operand(r9, ASR, kSmiShiftSize), SetCC);
+  __ b(cs, &convert_hole);  // Non-smi sets shifter carry.
 
   // Normal smi, convert to double and store.
-  __ SmiUntag(r9);
   if (vfp3_supported) {
     CpuFeatures::Scope scope(VFP3);
     __ vmov(s0, r9);
     __ vcvt_f64_s32(d0, s0);
     __ vstr(d0, r7, 0);
     __ add(r7, r7, Operand(8));
   } else {
     FloatingPointHelper::ConvertIntToDouble(masm,
                                             r9,
                                             FloatingPointHelper::kCoreRegisters,
                                             d0,
                                             r0,
                                             r1,
                                             lr,
                                             s0);
     __ Strd(r0, r1, MemOperand(r7, 8, PostIndex));
   }
   __ b(&entry);
 
   // Hole found, store the-hole NaN.
   __ bind(&convert_hole);
   if (FLAG_debug_code) {
+    // Restore a "smi-untagged" heap object.
+    __ SmiTag(r9);
+    __ orr(r9, r9, Operand(1));
     __ CompareRoot(r9, Heap::kTheHoleValueRootIndex);
     __ Assert(eq, "object found in smi-only array");
   }
   __ Strd(r4, r5, MemOperand(r7, 8, PostIndex));
 
   __ bind(&entry);
   __ cmp(r7, r6);
   __ b(lt, &loop);
 
   if (!vfp3_supported) __ Pop(r1, r0);
   __ pop(lr);
 }
 
 
 void ElementsTransitionGenerator::GenerateDoubleToObject(
     MacroAssembler* masm, Label* fail) {
   // ----------- S t a t e -------------
   //  -- r0    : value
   //  -- r1    : key
   //  -- r2    : receiver
   //  -- lr    : return address
   //  -- r3    : target map, scratch for subsequent call
   //  -- r4    : scratch (elements)
   // -----------------------------------
   Label entry, loop, convert_hole, gc_required;
 
   __ push(lr);
+  __ ldr(r4, FieldMemOperand(r2, JSObject::kElementsOffset));
   __ Push(r3, r2, r1, r0);
-
-  __ ldr(r4, FieldMemOperand(r2, JSObject::kElementsOffset));
   __ ldr(r5, FieldMemOperand(r4, FixedArray::kLengthOffset));
   // r4: source FixedDoubleArray
   // r5: number of elements (smi-tagged)
 
   // Allocate new FixedArray.
   __ mov(r0, Operand(FixedDoubleArray::kHeaderSize));
   __ add(r0, r0, Operand(r5, LSL, 1));
   __ AllocateInNewSpace(r0, r6, r7, r9, &gc_required, NO_ALLOCATION_FLAGS);
   // r6: destination FixedArray, not tagged as heap object
+  // Set destination FixedDoubleArray's length and map.
   __ LoadRoot(r9, Heap::kFixedArrayMapRootIndex);
+  __ str(r5, MemOperand(r6, FixedDoubleArray::kLengthOffset));
   __ str(r9, MemOperand(r6, HeapObject::kMapOffset));
-  // Set destination FixedDoubleArray's length.
-  __ str(r5, MemOperand(r6, FixedDoubleArray::kLengthOffset));
 
   // Prepare for conversion loop.
   __ add(r4, r4, Operand(FixedDoubleArray::kHeaderSize - kHeapObjectTag + 4));
   __ add(r3, r6, Operand(FixedArray::kHeaderSize));
   __ add(r6, r6, Operand(kHeapObjectTag));
   __ add(r5, r3, Operand(r5, LSL, 1));
   __ LoadRoot(r7, Heap::kTheHoleValueRootIndex);
   __ LoadRoot(r9, Heap::kHeapNumberMapRootIndex);
   // Using offsetted addresses in r4 to fully take advantage of post-indexing.
   // r3: begin of destination FixedArray element fields, not tagged
@@ skipping 81 matching lines @@
   __ b(eq, &check_sequential);
 
   // Dispatch on the indirect string shape: slice or cons.
   Label cons_string;
   __ tst(result, Operand(kSlicedNotConsMask));
   __ b(eq, &cons_string);
 
   // Handle slices.
   Label indirect_string_loaded;
   __ ldr(result, FieldMemOperand(string, SlicedString::kOffsetOffset));
+  __ ldr(string, FieldMemOperand(string, SlicedString::kParentOffset));
   __ add(index, index, Operand(result, ASR, kSmiTagSize));
-  __ ldr(string, FieldMemOperand(string, SlicedString::kParentOffset));
   __ jmp(&indirect_string_loaded);
 
   // Handle cons strings.
   // Check whether the right hand side is the empty string (i.e. if
   // this is really a flat string in a cons string). If that is not
   // the case we would rather go to the runtime system now to flatten
   // the string.
   __ bind(&cons_string);
   __ ldr(result, FieldMemOperand(string, ConsString::kSecondOffset));
-  __ LoadRoot(ip, Heap::kEmptyStringRootIndex);
-  __ cmp(result, ip);
-  __ b(ne, call_runtime);
+  __ CompareRoot(result, Heap::kEmptyStringRootIndex);
   // Get the first of the two strings and load its instance type.
+  __ b(ne, call_runtime);  // This branch refers to the previous CompareRoot.

[ulan, 2012/01/27 13:19:11]

This branch should be above the comment in line 343.

   __ ldr(string, FieldMemOperand(string, ConsString::kFirstOffset));
 
   __ bind(&indirect_string_loaded);
   __ ldr(result, FieldMemOperand(string, HeapObject::kMapOffset));
   __ ldrb(result, FieldMemOperand(result, Map::kInstanceTypeOffset));
 
   // Distinguish sequential and external strings. Only these two string
   // representations can reach here (slices and flat cons strings have been
   // reduced to the underlying sequential or external string).
   Label external_string, check_encoding;
@@ skipping 35 matching lines @@
   // Ascii string.
   __ ldrb(result, MemOperand(string, index));
   __ bind(&done);
 }
 
 #undef __
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_ARM