Unboxed load/store indexed of Float32x4

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 1106 matching lines @@
 
 CompileType LoadIndexedInstr::ComputeType() const {
   switch (class_id_) {
     case kArrayCid:
     case kImmutableArrayCid:
       return CompileType::Dynamic();
 
     case kTypedDataFloat32ArrayCid:
     case kTypedDataFloat64ArrayCid:
       return CompileType::FromCid(kDoubleCid);
+    case kTypedDataFloat32x4ArrayCid:
+      return CompileType::FromCid(kFloat32x4Cid);
 
     case kTypedDataInt8ArrayCid:
     case kTypedDataUint8ArrayCid:
     case kTypedDataUint8ClampedArrayCid:
     case kExternalTypedDataUint8ArrayCid:
     case kExternalTypedDataUint8ClampedArrayCid:
     case kTypedDataInt16ArrayCid:
     case kTypedDataUint16ArrayCid:
     case kOneByteStringCid:
     case kTwoByteStringCid:
@@ skipping 29 matching lines @@
     case kTwoByteStringCid:
       return kTagged;
     case kTypedDataInt32ArrayCid:
     case kTypedDataUint32ArrayCid:
       // Instruction can deoptimize if we optimistically assumed that the result
       // fits into Smi.
       return CanDeoptimize() ? kTagged : kUnboxedMint;
     case kTypedDataFloat32ArrayCid:
     case kTypedDataFloat64ArrayCid:
       return kUnboxedDouble;
+    case kTypedDataFloat32x4ArrayCid:
+      return kUnboxedFloat32x4;
     default:
       UNIMPLEMENTED();
       return kTagged;
   }
 }
 
 
 LocationSummary* LoadIndexedInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 2;
   const intptr_t kNumTemps = 0;
@@ skipping 37 matching lines @@
     ASSERT(this->array()->definition()->representation() == kTagged);
     element_address = index.IsRegister()
         ? FlowGraphCompiler::ElementAddressForRegIndex(
             class_id(), index_scale(), array, index.reg())
         : FlowGraphCompiler::ElementAddressForIntIndex(
             class_id(), index_scale(), array,
             Smi::Cast(index.constant()).Value());
   }
 
   if ((representation() == kUnboxedDouble) ||
-      (representation() == kUnboxedMint)) {
+      (representation() == kUnboxedMint) ||
+      (representation() == kUnboxedFloat32x4)) {
     XmmRegister result = locs()->out().fpu_reg();
     if ((index_scale() == 1) && index.IsRegister()) {
       __ SmiUntag(index.reg());
     }
     switch (class_id()) {
       case kTypedDataInt32ArrayCid:
         __ movss(result, element_address);
         __ pmovsxdq(result, result);
         break;
       case kTypedDataUint32ArrayCid:
         __ xorpd(result, result);
         __ movss(result, element_address);
         break;
       case kTypedDataFloat32ArrayCid:
         // Load single precision float and promote to double.
         __ movss(result, element_address);
         __ cvtss2sd(result, locs()->out().fpu_reg());
         break;
       case kTypedDataFloat64ArrayCid:
         __ movsd(result, element_address);
         break;
+      case kTypedDataFloat32x4ArrayCid:
+        __ movups(result, element_address);
+        break;
     }
     return;
   }
 
   Register result = locs()->out().reg();
   if ((index_scale() == 1) && index.IsRegister()) {
     __ SmiUntag(index.reg());
   }
   switch (class_id()) {
     case kTypedDataInt8ArrayCid:
@@ skipping 60 matching lines @@
     case kExternalTypedDataUint8ClampedArrayCid:
     case kTypedDataInt16ArrayCid:
     case kTypedDataUint16ArrayCid:
       return kTagged;
     case kTypedDataInt32ArrayCid:
     case kTypedDataUint32ArrayCid:
       return value()->IsSmiValue() ? kTagged : kUnboxedMint;
     case kTypedDataFloat32ArrayCid:
     case kTypedDataFloat64ArrayCid:
       return kUnboxedDouble;
+    case kTypedDataFloat32x4ArrayCid:
+      return kUnboxedFloat32x4;
     default:
       UNIMPLEMENTED();
       return kTagged;
   }
 }
 
 
 LocationSummary* StoreIndexedInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 3;
   const intptr_t kNumTemps = 0;
@@ skipping 42 matching lines @@
                       : Location::RequiresFpuRegister());
       break;
     case kTypedDataFloat32ArrayCid:
       // Need temp register for float-to-double conversion.
       locs->AddTemp(Location::RequiresFpuRegister());
       // Fall through.
     case kTypedDataFloat64ArrayCid:
       // TODO(srdjan): Support Float64 constants.
       locs->set_in(2, Location::RequiresFpuRegister());
       break;
+    case kTypedDataFloat32x4ArrayCid:
+      locs->set_in(2, Location::RequiresFpuRegister());
+      break;
     default:
       UNREACHABLE();
       return NULL;
   }
   return locs;
 }
 
 
 void StoreIndexedInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
   Register array = locs()->in(0).reg();
@@ skipping 96 matching lines @@
       break;
     case kTypedDataFloat32ArrayCid:
       // 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 kTypedDataFloat64ArrayCid:
       __ movsd(element_address, locs()->in(2).fpu_reg());
       break;
+    case kTypedDataFloat32x4ArrayCid:
+      __ movups(element_address, locs()->in(2).fpu_reg());
+      break;
     default:
       UNREACHABLE();
   }
 }
 
 
 LocationSummary* GuardFieldInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 1;
   LocationSummary* summary =
       new LocationSummary(kNumInputs, 0, LocationSummary::kNoCall);
@@ skipping 1158 matching lines @@
     __ jmp(&done);
     __ Bind(&is_smi);
     __ movl(temp, value);
     __ SmiUntag(temp);
     __ cvtsi2sd(result, temp);
     __ Bind(&done);
   }
 }
 
 
+LocationSummary* BoxFloat32x4Instr::MakeLocationSummary() 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)
+      : 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(instruction_->token_pos(),
+                           &label,
+                           PcDescriptors::kOther,
+                           locs);
+    if (EAX != locs->out().reg()) __ movl(locs->out().reg(), EAX);

[srdjan, 2013/04/10 15:33:23]

Use MoveRegister(locs->out().reg(), EAX), it emits code only if registers are
different.

[Cutch, 2013/04/10 16:36:13]

Done. I also made this change in the BoxDoubleSlowPath and BoxIntegerSlowPath.

+    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,
+                 out_reg);
+  __ Bind(slow_path->exit_label());
+  __ movups(FieldAddress(out_reg, Float32x4::value_offset()), value);
+}
+
+
+LocationSummary* UnboxFloat32x4Instr::MakeLocationSummary() const {
+  const intptr_t kNumInputs = 1;
+  const intptr_t kNumTemps = 0;
+  LocationSummary* summary =
+      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
+  summary->set_in(0, Location::RequiresRegister());
+  summary->set_out(Location::RequiresFpuRegister());
+  return summary;
+}
+
+
+void UnboxFloat32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  const intptr_t value_cid = value()->Type()->ToCid();
+  const Register value = locs()->in(0).reg();
+  const XmmRegister result = locs()->out().fpu_reg();
+
+  if (value_cid == kFloat32x4Cid) {
+    __ movups(result, FieldAddress(value, Float32x4::value_offset()));
+  } else {
+    Label* deopt = compiler->AddDeoptStub(deopt_id_, kDeoptCheckClass);
+    Register temp = locs()->temp(0).reg();
+    __ CompareClassId(value, kFloat32x4Cid, temp);
+    __ j(NOT_EQUAL, deopt);
+    __ movups(result, FieldAddress(value, Float32x4::value_offset()));
+  }
+}
+
+
 LocationSummary* BinaryDoubleOpInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 2;
   const intptr_t kNumTemps = 0;
   LocationSummary* summary =
       new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
   summary->set_in(0, Location::RequiresFpuRegister());
   summary->set_in(1, Location::RequiresFpuRegister());
   summary->set_out(Location::SameAsFirstInput());
   return summary;
 }
@@ skipping 1018 matching lines @@
                          PcDescriptors::kOther,
                          locs());
   __ Drop(2);  // Discard type arguments and receiver.
 }
 
 }  // namespace dart
 
 #undef __
 
 #endif  // defined TARGET_ARCH_IA32