Unbox/Box Float64x2 and inline typed array loads and stores

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 816 matching lines @@
     case kImmutableArrayCid:
       return CompileType::Dynamic();
 
     case kTypedDataFloat32ArrayCid:
     case kTypedDataFloat64ArrayCid:
       return CompileType::FromCid(kDoubleCid);
     case kTypedDataFloat32x4ArrayCid:
       return CompileType::FromCid(kFloat32x4Cid);
     case kTypedDataInt32x4ArrayCid:
       return CompileType::FromCid(kInt32x4Cid);
+    case kTypedDataFloat64x2ArrayCid:
+      return CompileType::FromCid(kFloat64x2Cid);
 
     case kTypedDataInt8ArrayCid:
     case kTypedDataUint8ArrayCid:
     case kTypedDataUint8ClampedArrayCid:
     case kExternalTypedDataUint8ArrayCid:
     case kExternalTypedDataUint8ClampedArrayCid:
     case kTypedDataInt16ArrayCid:
     case kTypedDataUint16ArrayCid:
     case kOneByteStringCid:
     case kTwoByteStringCid:
@@ skipping 33 matching lines @@
       // 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;
     case kTypedDataInt32x4ArrayCid:
       return kUnboxedInt32x4;
+    case kTypedDataFloat64x2ArrayCid:
+      return kUnboxedFloat64x2;
     default:
       UNIMPLEMENTED();
       return kTagged;
   }
 }
 
 
 LocationSummary* LoadIndexedInstr::MakeLocationSummary(bool opt) 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());
   if (CanBeImmediateIndex(index(), class_id())) {
     // CanBeImmediateIndex must return false for unsafe smis.
     locs->set_in(1, Location::Constant(index()->BoundConstant()));
   } else {
     // The index is either untagged (element size == 1) or a smi (for all
     // element sizes > 1).
     locs->set_in(1, (index_scale() == 1)
                         ? Location::WritableRegister()
                         : Location::RequiresRegister());
   }
   if ((representation() == kUnboxedDouble) ||
       (representation() == kUnboxedFloat32x4) ||
-      (representation() == kUnboxedInt32x4)) {
+      (representation() == kUnboxedInt32x4) ||
+      (representation() == kUnboxedFloat64x2)) {
     locs->set_out(Location::RequiresFpuRegister());
   } else {
     locs->set_out(Location::RequiresRegister());
   }
   return locs;
 }
 
 
 void LoadIndexedInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
   Register array = locs()->in(0).reg();
@@ skipping 12 matching lines @@
         ? 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() == kUnboxedFloat32x4) ||
-      (representation() == kUnboxedInt32x4)) {
+      (representation() == kUnboxedInt32x4) ||
+      (representation() == kUnboxedFloat64x2)) {
+    if (class_id_ == kTypedDataFloat64x2ArrayCid) {

[Florian Schneider, 2014/02/20 15:54:44]

Left-over debug printing?

[Cutch, 2014/02/20 19:33:11]

Done.

+      printf("class id match\n");
+      printf("representation = %d\n", representation());
+      printf("%d\n", locs()->out().kind());
+    }
     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 kTypedDataInt32x4ArrayCid:
       case kTypedDataFloat32x4ArrayCid:
+      case kTypedDataFloat64x2ArrayCid:
         __ movups(result, element_address);
         break;
     }
     return;
   }
 
   Register result = locs()->out().reg();
   if ((index_scale() == 1) && index.IsRegister()) {
     __ SmiUntag(index.reg());
   }
@@ skipping 67 matching lines @@
     case kTypedDataInt32ArrayCid:
     case kTypedDataUint32ArrayCid:
       return value()->IsSmiValue() ? kTagged : kUnboxedMint;
     case kTypedDataFloat32ArrayCid:
     case kTypedDataFloat64ArrayCid:
       return kUnboxedDouble;
     case kTypedDataFloat32x4ArrayCid:
       return kUnboxedFloat32x4;
     case kTypedDataInt32x4ArrayCid:
       return kUnboxedInt32x4;
+    case kTypedDataFloat64x2ArrayCid:
+      return kUnboxedFloat64x2;
     default:
       UNIMPLEMENTED();
       return kTagged;
   }
 }
 
 
 LocationSummary* StoreIndexedInstr::MakeLocationSummary(bool opt) const {
   const intptr_t kNumInputs = 3;
   const intptr_t kNumTemps = 0;
@@ skipping 42 matching lines @@
     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 kTypedDataInt32x4ArrayCid:
     case kTypedDataFloat32x4ArrayCid:
+    case kTypedDataFloat64x2ArrayCid:
       locs->set_in(2, Location::RequiresFpuRegister());
       break;
     default:
       UNREACHABLE();
       return NULL;
   }
   return locs;
 }
 
 
@@ skipping 101 matching lines @@
       // 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 kTypedDataInt32x4ArrayCid:
     case kTypedDataFloat32x4ArrayCid:
+    case kTypedDataFloat64x2ArrayCid:
       __ movups(element_address, locs()->in(2).fpu_reg());
       break;
     default:
       UNREACHABLE();
   }
 }
 
 
 LocationSummary* GuardFieldInstr::MakeLocationSummary(bool opt) const {
   const intptr_t kNumInputs = 1;
@@ skipping 736 matching lines @@
     compiler->RestoreLiveRegisters(locs);
 
     __ jmp(exit_label());
   }
 
  private:
   Instruction* instruction_;
 };
 
 
+class BoxFloat64x2SlowPath : public SlowPathCode {
+ public:
+  explicit BoxFloat64x2SlowPath(Instruction* instruction)
+      : instruction_(instruction) { }
+
+  virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
+    __ Comment("BoxFloat64x2SlowPath");
+    __ Bind(entry_label());
+    const Class& float64x2_class = compiler->float64x2_class();
+    const Code& stub =
+        Code::Handle(StubCode::GetAllocationStubForClass(float64x2_class));
+    const ExternalLabel label(float64x2_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 1027 matching lines @@
     Label* deopt = compiler->AddDeoptStub(deopt_id_, kDeoptCheckClass);
     __ testl(value, Immediate(kSmiTagMask));
     __ j(ZERO, deopt);
     __ CompareClassId(value, kFloat32x4Cid, temp);
     __ j(NOT_EQUAL, deopt);
   }
   __ movups(result, FieldAddress(value, Float32x4::value_offset()));
 }
 
 
+LocationSummary* BoxFloat64x2Instr::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 BoxFloat64x2Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  BoxFloat64x2SlowPath* slow_path = new BoxFloat64x2SlowPath(this);
+  compiler->AddSlowPathCode(slow_path);
+
+  Register out_reg = locs()->out().reg();
+  XmmRegister value = locs()->in(0).fpu_reg();
+
+  __ TryAllocate(compiler->float64x2_class(),
+                 slow_path->entry_label(),
+                 Assembler::kFarJump,
+                 out_reg,
+                 kNoRegister);
+  __ Bind(slow_path->exit_label());
+  __ movups(FieldAddress(out_reg, Float64x2::value_offset()), value);
+}
+
+
+LocationSummary* UnboxFloat64x2Instr::MakeLocationSummary(bool opt) const {
+  const intptr_t value_cid = value()->Type()->ToCid();
+  const intptr_t kNumInputs = 1;
+  const intptr_t kNumTemps = value_cid == kFloat64x2Cid ? 0 : 1;
+  LocationSummary* summary =
+      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
+  summary->set_in(0, Location::RequiresRegister());
+  if (kNumTemps > 0) {
+    ASSERT(kNumTemps == 1);
+    summary->set_temp(0, Location::RequiresRegister());
+  }
+  summary->set_out(Location::RequiresFpuRegister());
+  return summary;
+}
+
+
+void UnboxFloat64x2Instr::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 != kFloat64x2Cid) {
+    const Register temp = locs()->temp(0).reg();
+    Label* deopt = compiler->AddDeoptStub(deopt_id_, kDeoptCheckClass);
+    __ testl(value, Immediate(kSmiTagMask));
+    __ j(ZERO, deopt);
+    __ CompareClassId(value, kFloat64x2Cid, temp);
+    __ j(NOT_EQUAL, deopt);
+  }
+  __ movups(result, FieldAddress(value, Float64x2::value_offset()));
+}
+
+
 LocationSummary* BoxInt32x4Instr::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 2220 matching lines @@
                          PcDescriptors::kOther,
                          locs());
   __ Drop(2);  // Discard type arguments and receiver.
 }
 
 }  // namespace dart
 
 #undef __
 
 #endif  // defined TARGET_ARCH_IA32