Optimize loads and stores to Int32Array and Uint32Array.

runtime/vm/intermediate_language_ia32.cc

 // Copyright (c) 2013, the Dart project authors.  Please see the AUTHORS file
 // for details. All rights reserved. Use of this source code is governed by a
 // BSD-style license that can be found in the LICENSE file.
 
 #include "vm/globals.h"  // Needed here to get TARGET_ARCH_IA32.
 #if defined(TARGET_ARCH_IA32)
 
 #include "vm/intermediate_language.h"
 
 #include "lib/error.h"
@@ skipping 1073 matching lines @@
   Register result = locs()->out().reg();
   __ movl(result,
           Immediate(reinterpret_cast<uword>(Symbols::PredefinedAddress())));
   __ movl(result, Address(result,
                           char_code,
                           TIMES_HALF_WORD_SIZE,  // Char code is a smi.
                           Symbols::kNullCharCodeSymbolOffset * kWordSize));
 }
 
 
+intptr_t LoadIndexedInstr::ResultCid() const {
+  switch (class_id_) {
+    case kArrayCid:
+    case kImmutableArrayCid:
+      return kDynamicCid;
+    case kFloat32ArrayCid :
+    case kFloat64ArrayCid :
+      return kDoubleCid;
+    case kInt8ArrayCid:
+    case kUint8ArrayCid:
+    case kUint8ClampedArrayCid:
+    case kExternalUint8ArrayCid:
+    case kInt16ArrayCid:
+    case kUint16ArrayCid:
+    case kOneByteStringCid:
+    case kTwoByteStringCid:
+      return kSmiCid;
+    case kInt32ArrayCid:
+    case kUint32ArrayCid:
+      // Result can be smi or mint when boxed.
+      return kDynamicCid;
+    default:
+      UNIMPLEMENTED();
+      return kDynamicCid;
+  }
+}
+
+
+Representation LoadIndexedInstr::representation() const {
+  switch (class_id_) {
+    case kArrayCid:
+    case kImmutableArrayCid:
+    case kInt8ArrayCid:
+    case kUint8ArrayCid:
+    case kUint8ClampedArrayCid:
+    case kExternalUint8ArrayCid:
+    case kInt16ArrayCid:
+    case kUint16ArrayCid:
+    case kOneByteStringCid:
+    case kTwoByteStringCid:
+      return kTagged;
+    case kInt32ArrayCid:
+    case kUint32ArrayCid:
+      return kUnboxedMint;
+    case kFloat32ArrayCid :
+    case kFloat64ArrayCid :
+      return kUnboxedDouble;
+    default:
+      UNIMPLEMENTED();
+      return kTagged;
+  }
+}
+
+
 LocationSummary* LoadIndexedInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 2;
   const intptr_t kNumTemps = 0;
   LocationSummary* locs =
       new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
   locs->set_in(0, Location::RequiresRegister());
   // The smi index is either untagged and tagged again at the end of the
   // operation (element size == 1), or it is left smi tagged (for all element
   // sizes > 1).
   locs->set_in(1, CanBeImmediateIndex(index(), class_id())
@@ skipping 33 matching lines @@
     }
     return;
   }
 
   FieldAddress element_address = index.IsRegister()
       ? FlowGraphCompiler::ElementAddressForRegIndex(
           class_id(), array, index.reg())
       : FlowGraphCompiler::ElementAddressForIntIndex(
           class_id(), array, Smi::Cast(index.constant()).Value());
 
-  if (representation() == kUnboxedDouble) {
+  if (representation() == kUnboxedDouble ||

[Kevin Millikin (Google), 2013/01/21 15:16:01]

I guess you need more parens.

[Florian Schneider, 2013/01/21 16:03:25]

Done.

+      representation() == kUnboxedMint) {
     XmmRegister result = locs()->out().fpu_reg();
-    if (class_id() == kFloat32ArrayCid) {
-      // Load single precision float.
-      __ movss(result, element_address);
-      // Promote to double.
-      __ cvtss2sd(result, locs()->out().fpu_reg());
-    } else {
-      ASSERT(class_id() == kFloat64ArrayCid);
-      __ movsd(result, element_address);
+    switch (class_id()) {
+      case kInt32ArrayCid:
+        __ movss(result, element_address);
+        __ pmovsxdq(result, result);
+        break;
+      case kUint32ArrayCid:
+        __ xorpd(result, result);
+        __ movss(result, element_address);
+        break;
+      case kFloat32ArrayCid:
+        // Load single precision float and promote to double.
+        __ movss(result, element_address);
+        __ cvtss2sd(result, locs()->out().fpu_reg());
+        break;
+      case kFloat64ArrayCid:
+        __ movsd(result, element_address);
+        break;
     }
     return;
   }
 
   Register result = locs()->out().reg();
   switch (class_id()) {
     case kInt8ArrayCid:
     case kUint8ArrayCid:
     case kUint8ClampedArrayCid:
     case kOneByteStringCid:
@@ skipping 20 matching lines @@
       __ SmiTag(result);
       break;
     default:
       ASSERT((class_id() == kArrayCid) || (class_id() == kImmutableArrayCid));
       __ movl(result, element_address);
       break;
   }
 }
 
 
+Representation StoreIndexedInstr::RequiredInputRepresentation(
+    intptr_t idx) const {
+  if ((idx == 0) || (idx == 1)) return kTagged;
+  ASSERT(idx == 2);
+  switch (class_id_) {
+    case kArrayCid:
+    case kInt8ArrayCid:
+    case kUint8ArrayCid:
+    case kUint8ClampedArrayCid:
+    case kInt16ArrayCid:
+    case kUint16ArrayCid:
+      return kTagged;
+    case kInt32ArrayCid:
+    case kUint32ArrayCid:
+      return kUnboxedMint;
+    case kFloat32ArrayCid :
+    case kFloat64ArrayCid :
+      return kUnboxedDouble;
+    default:
+      UNIMPLEMENTED();
+      return kTagged;
+  }
+}
+
+
 LocationSummary* StoreIndexedInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 3;
   const intptr_t kNumTemps = 0;
   LocationSummary* locs =
       new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
   locs->set_in(0, Location::RequiresRegister());
   // The smi index is either untagged and tagged again at the end of the
   // operation (element size == 1), or it is left smi tagged (for all element
   // sizes > 1).
   locs->set_in(1, CanBeImmediateIndex(index(), class_id())
@@ skipping 14 matching lines @@
       break;
     case kInt16ArrayCid:
     case kUint16ArrayCid:
       // Writable register because the value must be untagged before storing.
       locs->set_in(2, Location::WritableRegister());
       break;
     case kFloat32ArrayCid:
       // Need temp register for float-to-double conversion.
       locs->AddTemp(Location::RequiresFpuRegister());
       // Fall through.
+    case kInt32ArrayCid:
+    case kUint32ArrayCid:
     case kFloat64ArrayCid:
       // TODO(srdjan): Support Float64 constants.
       locs->set_in(2, Location::RequiresFpuRegister());
       break;
     default:
       UNREACHABLE();
       return NULL;
   }
   return locs;
 }
@@ skipping 70 matching lines @@
         __ Bind(&store_value);
         __ movb(element_address, AL);
       }
       if (index.IsRegister()) {
         __ SmiTag(index.reg());  // Re-tag.
       }
       break;
     }
     case kInt16ArrayCid:
     case kUint16ArrayCid: {
-        Register value = locs()->in(2).reg();
-        __ SmiUntag(value);
-        __ movw(element_address, value);
-        break;
-      }
+      Register value = locs()->in(2).reg();
+      __ SmiUntag(value);
+      __ movw(element_address, value);
+      break;
+    }
+    case kInt32ArrayCid:
+    case kUint32ArrayCid:
+      __ movss(element_address, locs()->in(2).fpu_reg());
+      break;
     case kFloat32ArrayCid:
       // Convert to single precision.
       __ cvtsd2ss(locs()->temp(0).fpu_reg(), locs()->in(2).fpu_reg());
       // Store.
       __ movss(element_address, locs()->temp(0).fpu_reg());
       break;
     case kFloat64ArrayCid:
       __ movsd(element_address, locs()->in(2).fpu_reg());
       break;
     default:
@@ skipping 1879 matching lines @@
                          PcDescriptors::kOther,
                          locs());
   __ Drop(2);  // Discard type arguments and receiver.
 }
 
 }  // namespace dart
 
 #undef __
 
 #endif  // defined TARGET_ARCH_IA32