Support reusable boxes for Float32x4 fields

runtime/vm/intermediate_language_ia32.cc

Index: runtime/vm/intermediate_language_ia32.cc
diff --git a/runtime/vm/intermediate_language_ia32.cc b/runtime/vm/intermediate_language_ia32.cc
index c1b3e0435bfc0b34ac238c4fff8c7fa10052858e..323fbdb46135337579620880d43a47607f5437c4 100644
--- a/runtime/vm/intermediate_language_ia32.cc
+++ b/runtime/vm/intermediate_language_ia32.cc
@@ -15,6 +15,7 @@
 #include "vm/stack_frame.h"
 #include "vm/stub_code.h"
 #include "vm/symbols.h"
+#include "vm/il_printer.h"

[srdjan, 2014/02/05 22:20:09]

Please use alphabetic order.

[Cutch, 2014/02/05 23:16:02]

Removed.

 
 #define __ compiler->assembler()->
 
@@ -1590,36 +1591,101 @@ void GuardFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
 }
 
 
+static void PrintLocationSummary(const char* prefix, LocationSummary* locs) {
+  return;
+  char str[4000];
+  BufferFormatter f(str, sizeof(str));
+  locs->PrintTo(&f);
+  printf("%s: [%" Pd "] %s\n", prefix, locs->stack_bitmap()->Length(), str);
+}
+
+
+static void PrintStack(const char* prefix, FlowGraphCompiler* compiler) {
+  return;
+  printf("%s: StackSize() -> %" Pd "\n", prefix, compiler->StackSize());
+}
+
+
 class StoreInstanceFieldSlowPath : public SlowPathCode {
  public:
-  StoreInstanceFieldSlowPath(StoreInstanceFieldInstr* instruction,
-                             const Class& cls)
-      : instruction_(instruction), cls_(cls) { }
+  explicit StoreInstanceFieldSlowPath(StoreInstanceFieldInstr* instruction)
+      : instruction_(instruction) { }
 
   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);
+    {
+      __ Bind(double_entry_label());
+      const Class& cls = compiler->double_class();
+      const Code& stub =
+          Code::Handle(StubCode::GetAllocationStubForClass(cls));
+      const ExternalLabel label(cls.ToCString(), stub.EntryPoint());
+
+      LocationSummary* locs = instruction_->locs();
+      PrintLocationSummary("A", locs);
+      PrintStack("A", compiler);
+      locs->live_registers()->Remove(locs->out());
+      PrintLocationSummary("B", locs);
+      PrintStack("B", compiler);
+      compiler->SaveLiveRegisters(locs);
+      PrintLocationSummary("C", locs);
+      PrintStack("C", compiler);
+      compiler->GenerateCall(Scanner::kNoSourcePos,  // No token position.
+                             &label,
+                             PcDescriptors::kOther,
+                             locs);
+      PrintLocationSummary("D", locs);
+      PrintStack("D", compiler);
+      __ MoveRegister(locs->temp(0).reg(), EAX);
+      compiler->RestoreLiveRegisters(locs);
+      PrintLocationSummary("E", locs);
+      PrintStack("E", compiler);
+      __ jmp(double_exit_label());
+    }
+    {
+      __ Bind(float32x4_entry_label());
+      const Class& cls = compiler->float32x4_class();
+      const Code& stub =
+          Code::Handle(StubCode::GetAllocationStubForClass(cls));
+      const ExternalLabel label(cls.ToCString(), stub.EntryPoint());
+      LocationSummary* locs = instruction_->locs();
+      PrintLocationSummary("F", locs);
+      PrintStack("F", compiler);
+      locs->live_registers()->Remove(locs->out());
+      PrintLocationSummary("G", locs);
+      PrintStack("G", compiler);
+      compiler->SaveLiveRegisters(locs);
+      PrintLocationSummary("H", locs);
+      PrintStack("H", compiler);
+      compiler->GenerateCall(Scanner::kNoSourcePos,  // No token position.
+                             &label,
+                             PcDescriptors::kOther,
+                             locs);
+      PrintLocationSummary("I", locs);
+      PrintStack("I", compiler);
+      __ MoveRegister(locs->temp(0).reg(), EAX);
+      compiler->RestoreLiveRegisters(locs);
+      PrintLocationSummary("J", locs);
+      PrintStack("J", compiler);
+      __ jmp(float32x4_exit_label());
+    }
+  }
 
-    __ jmp(exit_label());
+  Label* double_entry_label() {
+    // Use default SlowPathCode label for double.
+    return entry_label();
+  }
+  Label* double_exit_label() {
+    // Use default SlowPathCode label for double.
+    return exit_label();
   }
 
+  Label* float32x4_entry_label() { return &float32x4_entry_label_; }
+  Label* float32x4_exit_label() { return &float32x4_exit_label_; }
+
  private:
+  Label float32x4_entry_label_;
+  Label float32x4_exit_label_;
   StoreInstanceFieldInstr* instruction_;
-  const Class& cls_;
 };
 
 
@@ -1666,25 +1732,34 @@ void StoreInstanceFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
     const intptr_t cid = field().UnboxedFieldCid();
 
     if (is_initialization_) {
+      StoreInstanceFieldSlowPath* slow_path =
+          new StoreInstanceFieldSlowPath(this);
+      compiler->AddSlowPathCode(slow_path);
+
       const Class* cls = NULL;
+      Label* entry_label = NULL;
+      Label* exit_label = NULL;
       switch (cid) {
         case kDoubleCid:
           cls = &compiler->double_class();
+          entry_label = slow_path->double_entry_label();
+          exit_label = slow_path->double_exit_label();
+          break;
+        case kFloat32x4Cid:
+          cls = &compiler->float32x4_class();
+          entry_label = slow_path->float32x4_entry_label();
+          exit_label = slow_path->float32x4_exit_label();
           break;
-        // TODO(johnmccutchan): Add kFloat32x4Cid here.
         default:
           UNREACHABLE();
       }
-      StoreInstanceFieldSlowPath* slow_path =
-          new StoreInstanceFieldSlowPath(this, *cls);
-      compiler->AddSlowPathCode(slow_path);
 
       __ TryAllocate(*cls,
-                     slow_path->entry_label(),
+                     entry_label,
                      Assembler::kFarJump,
                      temp,
                      temp2);
-      __ Bind(slow_path->exit_label());
+      __ Bind(exit_label);
       __ movl(temp2, temp);
       __ StoreIntoObject(instance_reg,
                          FieldAddress(instance_reg, field().Offset()),
@@ -1694,8 +1769,12 @@ void StoreInstanceFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
     }
     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();
@@ -1710,8 +1789,8 @@ void StoreInstanceFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
     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()));
 
@@ -1727,41 +1806,80 @@ void StoreInstanceFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
             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());
+      new StoreInstanceFieldSlowPath(this);
     compiler->AddSlowPathCode(slow_path);
 
-    if (!compiler->is_optimizing()) {
-      locs()->live_registers()->Add(locs()->in(0));
-      locs()->live_registers()->Add(locs()->in(1));
+    {
+      __ Bind(&store_double);
+      Label copy_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);
+
+      if (!compiler->is_optimizing()) {
+        locs()->live_registers()->Add(locs()->in(0));
+        locs()->live_registers()->Add(locs()->in(1));
+      }
+
+      __ TryAllocate(compiler->double_class(),
+                     slow_path->double_entry_label(),
+                     Assembler::kFarJump,
+                     temp,
+                     temp2);
+      __ Bind(slow_path->double_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);
     }
 
-    __ 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_float32x4);
+      Label copy_float32x4;
+
+      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);
+
+      if (!compiler->is_optimizing()) {
+        locs()->live_registers()->Add(locs()->in(0));
+        locs()->live_registers()->Add(locs()->in(1));
+      }
+
+      __ TryAllocate(compiler->float32x4_class(),
+                     slow_path->float32x4_entry_label(),
+                     Assembler::kFarJump,
+                     temp,
+                     temp2);
+      __ Bind(slow_path->float32x4_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(&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);
   }
 
@@ -1906,34 +2024,70 @@ void AllocateObjectWithBoundsCheckInstr::EmitNativeCode(
 }
 
 
-class BoxDoubleSlowPath : public SlowPathCode {
+class LoadFieldSlowPath : public SlowPathCode {
  public:
-  explicit BoxDoubleSlowPath(Instruction* instruction)
+  explicit LoadFieldSlowPath(Instruction* instruction)
       : instruction_(instruction) { }
 
   virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
-    __ Comment("BoxDoubleSlowPath");
-    __ Bind(entry_label());
-    const Class& double_class = compiler->double_class();
-    const Code& stub =
-        Code::Handle(StubCode::GetAllocationStubForClass(double_class));
-    const ExternalLabel label(double_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);
+    __ Comment("LoadFieldSlowPath");
+    {
+      __ Bind(double_entry_label());
+      const Class& double_class = compiler->double_class();
+      const Code& stub =
+          Code::Handle(StubCode::GetAllocationStubForClass(double_class));
+      const ExternalLabel label(double_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(double_exit_label());
+    }
+    {
+      __ Bind(float32x4_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(float32x4_exit_label());
+    }
+  }
 
-    __ jmp(exit_label());
+  Label* double_entry_label() {
+    // Use default SlowPathCode label for double.
+    return entry_label();
+  }
+  Label* double_exit_label() {
+    // Use default SlowPathCode label for double.
+    return exit_label();
   }
 
+  Label* float32x4_entry_label() { return &float32x4_entry_label_; }
+  Label* float32x4_exit_label() { return &float32x4_exit_label_; }
+
  private:
+  Label float32x4_entry_label_;
+  Label float32x4_exit_label_;
   Instruction* instruction_;
 };
 
@@ -1971,9 +2125,13 @@ void LoadFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
     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();
     }
@@ -1985,9 +2143,13 @@ void LoadFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
   if (IsPotentialUnboxedLoad()) {
     Register temp = locs()->temp(1).reg();
     XmmRegister value = locs()->temp(0).fpu_reg();
+    LoadFieldSlowPath* slow_path = new LoadFieldSlowPath(this);
+    compiler->AddSlowPathCode(slow_path);
 
     Label load_pointer;
     Label load_double;
+    Label load_float32x4;
+
     __ LoadObject(result, Field::ZoneHandle(field()->raw()));
 
     FieldAddress field_cid_operand(result, Field::guarded_cid_offset());
@@ -1999,29 +2161,49 @@ void LoadFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
     __ 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);
+    {
+      __ Bind(&load_double);
+
+      if (!compiler->is_optimizing()) {
+        locs()->live_registers()->Add(locs()->in(0));
+      }
 
-    if (!compiler->is_optimizing()) {
-      locs()->live_registers()->Add(locs()->in(0));
+      __ TryAllocate(compiler->double_class(),
+                     slow_path->double_entry_label(),
+                     Assembler::kFarJump,
+                     result,
+                     temp);
+      __ Bind(slow_path->double_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);
     }
 
-    __ 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);
 
-    // TODO(johnmccutchan): Add Float32x4 path here.
+      if (!compiler->is_optimizing()) {
+        locs()->live_registers()->Add(locs()->in(0));
+      }
+
+      __ TryAllocate(compiler->float32x4_class(),
+                     slow_path->float32x4_entry_label(),
+                     Assembler::kFarJump,
+                     result,
+                     temp);
+      __ Bind(slow_path->float32x4_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);
   }
@@ -2877,6 +3059,38 @@ void CheckEitherNonSmiInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
 }
 
 
+class BoxDoubleSlowPath : public SlowPathCode {
+ public:
+  explicit BoxDoubleSlowPath(BoxDoubleInstr* instruction)
+      : instruction_(instruction) { }
+
+  virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
+    __ Comment("BoxDoubleSlowPath");
+    __ Bind(entry_label());
+    const Class& double_class = compiler->double_class();
+    const Code& stub =
+        Code::Handle(StubCode::GetAllocationStubForClass(double_class));
+    const ExternalLabel label(double_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:
+  BoxDoubleInstr* instruction_;
+};
+
+
 LocationSummary* BoxDoubleInstr::MakeLocationSummary(bool opt) const {
   const intptr_t kNumInputs = 1;
   const intptr_t kNumTemps = 0;
@@ -2953,19 +3167,6 @@ void UnboxDoubleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
 }
 
 
-LocationSummary* BoxFloat32x4Instr::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;
-}
-
-
 class BoxFloat32x4SlowPath : public SlowPathCode {
  public:
   explicit BoxFloat32x4SlowPath(BoxFloat32x4Instr* instruction)
@@ -2998,6 +3199,19 @@ class BoxFloat32x4SlowPath : public SlowPathCode {
 };
 
 
+LocationSummary* BoxFloat32x4Instr::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;
+}
+
+
 void BoxFloat32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
   BoxFloat32x4SlowPath* slow_path = new BoxFloat32x4SlowPath(this);
   compiler->AddSlowPathCode(slow_path);