Support reusable boxes for Float32x4 fields

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 "vm/dart_entry.h"
@@ skipping 1587 matching lines @@
 
   virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
     __ Comment("StoreInstanceFieldSlowPath");
     __ Bind(entry_label());
     const Code& stub =
         Code::Handle(StubCode::GetAllocationStubForClass(cls_));
     const ExternalLabel label(cls_.ToCString(), stub.EntryPoint());
 
     LocationSummary* locs = instruction_->locs();
     locs->live_registers()->Remove(locs->out());
-
     compiler->SaveLiveRegisters(locs);
     compiler->GenerateCall(Scanner::kNoSourcePos,  // No token position.
                            &label,
                            PcDescriptors::kOther,
                            locs);
     __ MoveRegister(locs->temp(0).reg(), EAX);
     compiler->RestoreLiveRegisters(locs);
-
     __ jmp(exit_label());
   }
 
  private:
   StoreInstanceFieldInstr* instruction_;
   const Class& cls_;
 };
 
 
 LocationSummary* StoreInstanceFieldInstr::MakeLocationSummary(bool opt) const {
@@ skipping 37 matching lines @@
     Register temp = locs()->temp(0).reg();
     Register temp2 = locs()->temp(1).reg();
     const intptr_t cid = field().UnboxedFieldCid();
 
     if (is_initialization_) {
       const Class* cls = NULL;
       switch (cid) {
         case kDoubleCid:
           cls = &compiler->double_class();
           break;
-        // TODO(johnmccutchan): Add kFloat32x4Cid here.
+        case kFloat32x4Cid:
+          cls = &compiler->float32x4_class();
+          break;
         default:
           UNREACHABLE();
       }
+
       StoreInstanceFieldSlowPath* slow_path =
           new StoreInstanceFieldSlowPath(this, *cls);
       compiler->AddSlowPathCode(slow_path);
 
       __ TryAllocate(*cls,
                      slow_path->entry_label(),
                      Assembler::kFarJump,
                      temp,
                      temp2);
       __ Bind(slow_path->exit_label());
       __ movl(temp2, temp);
       __ StoreIntoObject(instance_reg,
                          FieldAddress(instance_reg, field().Offset()),
                          temp2);
     } else {
       __ movl(temp, FieldAddress(instance_reg, field().Offset()));
     }
     switch (cid) {
       case kDoubleCid:
-      __ movsd(FieldAddress(temp, Double::value_offset()), value);
-      // TODO(johnmccutchan): Add kFloat32x4Cid here.
+        __ Comment("UnboxedDoubleStoreInstanceFieldInstr");
+        __ movsd(FieldAddress(temp, Double::value_offset()), value);
+      break;
+      case kFloat32x4Cid:
+        __ Comment("UnboxedFloat32x4StoreInstanceFieldInstr");
+        __ movups(FieldAddress(temp, Float32x4::value_offset()), value);
       break;
       default:
         UNREACHABLE();
     }
     return;
   }
 
   if (IsPotentialUnboxedStore()) {
     Register value_reg = locs()->in(1).reg();
     Register temp = locs()->temp(0).reg();
     Register temp2 = locs()->temp(1).reg();
     FpuRegister fpu_temp = locs()->temp(2).fpu_reg();
 
     Label store_pointer;
-    Label copy_double;
     Label store_double;
+    Label store_float32x4;
 
     __ LoadObject(temp, Field::ZoneHandle(field().raw()));
 
     __ cmpl(FieldAddress(temp, Field::is_nullable_offset()),
             Immediate(kNullCid));
     __ j(EQUAL, &store_pointer);
 
     __ movzxb(temp2, FieldAddress(temp, Field::kind_bits_offset()));
     __ testl(temp2, Immediate(1 << Field::kUnboxingCandidateBit));
     __ j(ZERO, &store_pointer);
 
     __ cmpl(FieldAddress(temp, Field::guarded_cid_offset()),
             Immediate(kDoubleCid));
     __ j(EQUAL, &store_double);
 
+    __ cmpl(FieldAddress(temp, Field::guarded_cid_offset()),
+            Immediate(kFloat32x4Cid));
+    __ j(EQUAL, &store_float32x4);
+
     // Fall through.
     __ jmp(&store_pointer);
 
-    __ Bind(&store_double);
-
-    const Immediate& raw_null =
-        Immediate(reinterpret_cast<intptr_t>(Object::null()));
-    __ movl(temp, FieldAddress(instance_reg, field().Offset()));
-    __ cmpl(temp, raw_null);
-    __ j(NOT_EQUAL, &copy_double);
-
-    StoreInstanceFieldSlowPath* slow_path =
-        new StoreInstanceFieldSlowPath(this, compiler->double_class());
-    compiler->AddSlowPathCode(slow_path);
 
     if (!compiler->is_optimizing()) {
       locs()->live_registers()->Add(locs()->in(0));
       locs()->live_registers()->Add(locs()->in(1));
     }
 
-    __ TryAllocate(compiler->double_class(),
-                   slow_path->entry_label(),
-                   Assembler::kFarJump,
-                   temp,
-                   temp2);
-    __ Bind(slow_path->exit_label());
-    __ movl(temp2, temp);
-    __ StoreIntoObject(instance_reg,
-                       FieldAddress(instance_reg, field().Offset()),
-                       temp2);
+    {
+      __ Bind(&store_double);
+      Label copy_double;
 
-    __ Bind(&copy_double);
-    __ movsd(fpu_temp, FieldAddress(value_reg, Double::value_offset()));
-    __ movsd(FieldAddress(temp, Double::value_offset()), fpu_temp);
-    __ jmp(&skip_store);
+      StoreInstanceFieldSlowPath* slow_path =
+          new StoreInstanceFieldSlowPath(this, compiler->double_class());
+      compiler->AddSlowPathCode(slow_path);
+
+      const Immediate& raw_null =
+          Immediate(reinterpret_cast<intptr_t>(Object::null()));
+      __ movl(temp, FieldAddress(instance_reg, field().Offset()));
+      __ cmpl(temp, raw_null);
+      __ j(NOT_EQUAL, &copy_double);
+
+      __ TryAllocate(compiler->double_class(),
+                     slow_path->entry_label(),
+                     Assembler::kFarJump,
+                     temp,
+                     temp2);
+      __ Bind(slow_path->exit_label());
+      __ movl(temp2, temp);
+      __ StoreIntoObject(instance_reg,
+                         FieldAddress(instance_reg, field().Offset()),
+                         temp2);
+
+      __ Bind(&copy_double);
+      __ movsd(fpu_temp, FieldAddress(value_reg, Double::value_offset()));
+      __ movsd(FieldAddress(temp, Double::value_offset()), fpu_temp);
+      __ jmp(&skip_store);
+    }
+
+    {
+      __ Bind(&store_float32x4);
+      Label copy_float32x4;
+
+      StoreInstanceFieldSlowPath* slow_path =
+          new StoreInstanceFieldSlowPath(this, compiler->float32x4_class());
+      compiler->AddSlowPathCode(slow_path);
+
+      const Immediate& raw_null =
+          Immediate(reinterpret_cast<intptr_t>(Object::null()));
+      __ movl(temp, FieldAddress(instance_reg, field().Offset()));
+      __ cmpl(temp, raw_null);
+      __ j(NOT_EQUAL, &copy_float32x4);
+
+      __ TryAllocate(compiler->float32x4_class(),
+                     slow_path->entry_label(),
+                     Assembler::kFarJump,
+                     temp,
+                     temp2);
+      __ Bind(slow_path->exit_label());
+      __ movl(temp2, temp);
+      __ StoreIntoObject(instance_reg,
+                         FieldAddress(instance_reg, field().Offset()),
+                         temp2);
+
+      __ Bind(&copy_float32x4);
+      __ movups(fpu_temp, FieldAddress(value_reg, Float32x4::value_offset()));
+      __ movups(FieldAddress(temp, Float32x4::value_offset()), fpu_temp);
+      __ jmp(&skip_store);
+    }
     __ Bind(&store_pointer);
   }
 
   if (ShouldEmitStoreBarrier()) {
     Register value_reg = locs()->in(1).reg();
     __ StoreIntoObject(instance_reg,
                        FieldAddress(instance_reg, field().Offset()),
                        value_reg,
                        CanValueBeSmi());
   } else {
@@ skipping 156 matching lines @@
     compiler->RestoreLiveRegisters(locs);
 
     __ jmp(exit_label());
   }
 
  private:
   Instruction* instruction_;
 };
 
 
+class BoxFloat32x4SlowPath : public SlowPathCode {
+ public:
+  explicit BoxFloat32x4SlowPath(Instruction* instruction)
+      : instruction_(instruction) { }
+
+  virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
+    __ Comment("BoxFloat32x4SlowPath");
+    __ Bind(entry_label());
+    const Class& float32x4_class = compiler->float32x4_class();
+    const Code& stub =
+        Code::Handle(StubCode::GetAllocationStubForClass(float32x4_class));
+    const ExternalLabel label(float32x4_class.ToCString(), stub.EntryPoint());
+
+    LocationSummary* locs = instruction_->locs();
+    locs->live_registers()->Remove(locs->out());
+
+    compiler->SaveLiveRegisters(locs);
+    compiler->GenerateCall(Scanner::kNoSourcePos,  // No token position.
+                           &label,
+                           PcDescriptors::kOther,
+                           locs);
+    __ MoveRegister(locs->out().reg(), EAX);
+    compiler->RestoreLiveRegisters(locs);
+
+    __ jmp(exit_label());
+  }
+
+ private:
+  Instruction* instruction_;
+};
+
+
+
 LocationSummary* LoadFieldInstr::MakeLocationSummary(bool opt) const {
   const intptr_t kNumInputs = 1;
   const intptr_t kNumTemps = 0;
   LocationSummary* locs =
       new LocationSummary(
           kNumInputs, kNumTemps,
           (opt && !IsPotentialUnboxedLoad())
           ? LocationSummary::kNoCall
           : LocationSummary::kCallOnSlowPath);
 
@@ skipping 13 matching lines @@
 
 void LoadFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
   Register instance_reg = locs()->in(0).reg();
   if (IsUnboxedLoad() && compiler->is_optimizing()) {
     XmmRegister result = locs()->out().fpu_reg();
     Register temp = locs()->temp(0).reg();
     __ movl(temp, FieldAddress(instance_reg, offset_in_bytes()));
     const intptr_t cid = field()->UnboxedFieldCid();
     switch (cid) {
       case kDoubleCid:
+        __ Comment("UnboxedDoubleLoadFieldInstr");
         __ movsd(result, FieldAddress(temp, Double::value_offset()));
         break;
-      // TODO(johnmccutchan): Add Float32x4 path here.
+      case kFloat32x4Cid:
+        __ Comment("UnboxedFloat32x4LoadFieldInstr");
+        __ movups(result, FieldAddress(temp, Float32x4::value_offset()));
+        break;
       default:
         UNREACHABLE();
     }
     return;
   }
 
   Label done;
   Register result = locs()->out().reg();
   if (IsPotentialUnboxedLoad()) {
     Register temp = locs()->temp(1).reg();
     XmmRegister value = locs()->temp(0).fpu_reg();
 
+
     Label load_pointer;
     Label load_double;
+    Label load_float32x4;
+
     __ LoadObject(result, Field::ZoneHandle(field()->raw()));
 
     FieldAddress field_cid_operand(result, Field::guarded_cid_offset());
     FieldAddress field_nullability_operand(result, Field::is_nullable_offset());
 
     __ cmpl(field_nullability_operand, Immediate(kNullCid));
     __ j(EQUAL, &load_pointer);
 
     __ cmpl(field_cid_operand, Immediate(kDoubleCid));
     __ j(EQUAL, &load_double);
 
+    __ cmpl(field_cid_operand, Immediate(kFloat32x4Cid));
+    __ j(EQUAL, &load_float32x4);
+
     // Fall through.
     __ jmp(&load_pointer);
 
-    __ Bind(&load_double);
-    BoxDoubleSlowPath* slow_path = new BoxDoubleSlowPath(this);
-    compiler->AddSlowPathCode(slow_path);
-
     if (!compiler->is_optimizing()) {
       locs()->live_registers()->Add(locs()->in(0));
     }
 
-    __ TryAllocate(compiler->double_class(),
-                   slow_path->entry_label(),
-                   Assembler::kFarJump,
-                   result,
-                   temp);
-    __ Bind(slow_path->exit_label());
-    __ movl(temp, FieldAddress(instance_reg, offset_in_bytes()));
-    __ movsd(value, FieldAddress(temp, Double::value_offset()));
-    __ movsd(FieldAddress(result, Double::value_offset()), value);
-    __ jmp(&done);
+    {
+      __ Bind(&load_double);
+      BoxDoubleSlowPath* slow_path = new BoxDoubleSlowPath(this);
+      compiler->AddSlowPathCode(slow_path);
 
-    // TODO(johnmccutchan): Add Float32x4 path here.
+      __ TryAllocate(compiler->double_class(),
+                     slow_path->entry_label(),
+                     Assembler::kFarJump,
+                     result,
+                     temp);
+      __ Bind(slow_path->exit_label());
+      __ movl(temp, FieldAddress(instance_reg, offset_in_bytes()));
+      __ movsd(value, FieldAddress(temp, Double::value_offset()));
+      __ movsd(FieldAddress(result, Double::value_offset()), value);
+      __ jmp(&done);
+    }
+
+    {
+      __ Bind(&load_float32x4);
+
+      BoxFloat32x4SlowPath* slow_path = new BoxFloat32x4SlowPath(this);
+      compiler->AddSlowPathCode(slow_path);
+
+      __ TryAllocate(compiler->float32x4_class(),
+                     slow_path->entry_label(),
+                     Assembler::kFarJump,
+                     result,
+                     temp);
+      __ Bind(slow_path->exit_label());
+      __ movl(temp, FieldAddress(instance_reg, offset_in_bytes()));
+      __ movups(value, FieldAddress(temp, Float32x4::value_offset()));
+      __ movups(FieldAddress(result, Float32x4::value_offset()), value);
+      __ jmp(&done);
+    }
 
     __ Bind(&load_pointer);
   }
   __ movl(result, FieldAddress(instance_reg, offset_in_bytes()));
   __ Bind(&done);
 }
 
 
 LocationSummary* InstantiateTypeInstr::MakeLocationSummary(bool opt) const {
   const intptr_t kNumInputs = 1;
@@ skipping 835 matching lines @@
   } else {
     Register temp = locs()->temp(0).reg();
     __ movl(temp, left);
     __ orl(temp, right);
     __ testl(temp, Immediate(kSmiTagMask));
   }
   __ j(ZERO, deopt);
 }
 
 
+
+
+
 LocationSummary* BoxDoubleInstr::MakeLocationSummary(bool opt) const {
   const intptr_t kNumInputs = 1;
   const intptr_t kNumTemps = 0;
   LocationSummary* summary =
       new LocationSummary(kNumInputs,
                           kNumTemps,
                           LocationSummary::kCallOnSlowPath);
   summary->set_in(0, Location::RequiresFpuRegister());
   summary->set_out(Location::RequiresRegister());
   return summary;
@@ skipping 69 matching lines @@
   LocationSummary* summary =
       new LocationSummary(kNumInputs,
                           kNumTemps,
                           LocationSummary::kCallOnSlowPath);
   summary->set_in(0, Location::RequiresFpuRegister());
   summary->set_out(Location::RequiresRegister());
   return summary;
 }
 
 
-class BoxFloat32x4SlowPath : public SlowPathCode {
- public:
-  explicit BoxFloat32x4SlowPath(BoxFloat32x4Instr* instruction)
-      : instruction_(instruction) { }
-
-  virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
-    __ Comment("BoxFloat32x4SlowPath");
-    __ Bind(entry_label());
-    const Class& float32x4_class = compiler->float32x4_class();
-    const Code& stub =
-        Code::Handle(StubCode::GetAllocationStubForClass(float32x4_class));
-    const ExternalLabel label(float32x4_class.ToCString(), stub.EntryPoint());
-
-    LocationSummary* locs = instruction_->locs();
-    locs->live_registers()->Remove(locs->out());
-
-    compiler->SaveLiveRegisters(locs);
-    compiler->GenerateCall(Scanner::kNoSourcePos,  // No token position.
-                           &label,
-                           PcDescriptors::kOther,
-                           locs);
-    __ MoveRegister(locs->out().reg(), EAX);
-    compiler->RestoreLiveRegisters(locs);
-
-    __ jmp(exit_label());
-  }
-
- private:
-  BoxFloat32x4Instr* instruction_;
-};
-
-
 void BoxFloat32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
   BoxFloat32x4SlowPath* slow_path = new BoxFloat32x4SlowPath(this);
   compiler->AddSlowPathCode(slow_path);
 
   Register out_reg = locs()->out().reg();
   XmmRegister value = locs()->in(0).fpu_reg();
 
   __ TryAllocate(compiler->float32x4_class(),
                  slow_path->entry_label(),
                  Assembler::kFarJump,
@@ skipping 2270 matching lines @@
                          PcDescriptors::kOther,
                          locs());
   __ Drop(2);  // Discard type arguments and receiver.
 }
 
 }  // namespace dart
 
 #undef __
 
 #endif  // defined TARGET_ARCH_IA32