Unbox/Box Float64x2 and inline typed array loads and stores

runtime/vm/intermediate_language_arm.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_ARM.
 #if defined(TARGET_ARCH_ARM)
 
 #include "vm/intermediate_language.h"
 
 #include "vm/dart_entry.h"
@@ skipping 811 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 kTypedDataInt32x4ArrayCid:
       return kUnboxedInt32x4;
     case kTypedDataFloat32x4ArrayCid:
       return kUnboxedFloat32x4;
+    case kTypedDataFloat64x2ArrayCid:
+      return kUnboxedFloat64x2;
     default:
       UNREACHABLE();
       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());
   // The smi index is either untagged (element size == 1), or it is left smi
   // tagged (for all element sizes > 1).
   // TODO(regis): Revisit and see if the index can be immediate.
   locs->set_in(1, Location::WritableRegister());
-  if ((representation() == kUnboxedDouble) ||
+  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 29 matching lines @@
       UNREACHABLE();
   }
 
   if (!IsExternal()) {
     ASSERT(this->array()->definition()->representation() == kTagged);
     __ AddImmediate(index.reg(),
         FlowGraphCompiler::DataOffsetFor(class_id()) - kHeapObjectTag);
   }
   element_address = Address(array, index.reg(), LSL, 0);
 
-  if ((representation() == kUnboxedDouble) ||
-      (representation() == kUnboxedMint) ||
+  if ((representation() == kUnboxedDouble)    ||
+      (representation() == kUnboxedMint)      ||
       (representation() == kUnboxedFloat32x4) ||
-      (representation() == kUnboxedInt32x4)) {
+      (representation() == kUnboxedInt32x4)   ||
+      (representation() == kUnboxedFloat64x2)) {
     QRegister result = locs()->out().fpu_reg();
     DRegister dresult0 = EvenDRegisterOf(result);
     DRegister dresult1 = OddDRegisterOf(result);
     switch (class_id()) {
       case kTypedDataInt32ArrayCid:
         UNIMPLEMENTED();
         break;
       case kTypedDataUint32ArrayCid:
         UNIMPLEMENTED();
         break;
       case kTypedDataFloat32ArrayCid:
         // Load single precision float and promote to double.
         // vldrs does not support indexed addressing.
         __ add(index.reg(), index.reg(), ShifterOperand(array));
         element_address = Address(index.reg(), 0);
         __ vldrs(STMP, element_address);
         __ vcvtds(dresult0, STMP);
         break;
       case kTypedDataFloat64ArrayCid:
         // vldrd does not support indexed addressing.
         __ add(index.reg(), index.reg(), ShifterOperand(array));
         element_address = Address(index.reg(), 0);
         __ vldrd(dresult0, element_address);
         break;
+      case kTypedDataFloat64x2ArrayCid:
       case kTypedDataInt32x4ArrayCid:
       case kTypedDataFloat32x4ArrayCid:
         __ add(index.reg(), index.reg(), ShifterOperand(array));
         __ LoadDFromOffset(dresult0, index.reg(), 0);

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

Isn't there a Load-quad-word instruction on ARM? Could you maybe use vldmd
load-multiple instruction?

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

zra says there is but it can only be used if there is no offset, I think we can
use it some places but not others. I've added a TODO to look at this.

         __ LoadDFromOffset(dresult1, index.reg(), 2*kWordSize);
         break;
     }
     return;
   }
 
   Register result = locs()->out().reg();
   switch (class_id()) {
     case kTypedDataInt8ArrayCid:
       ASSERT(index_scale() == 1);
@@ skipping 64 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:
       UNREACHABLE();
       return kTagged;
   }
 }
 
 
 LocationSummary* StoreIndexedInstr::MakeLocationSummary(bool opt) const {
   const intptr_t kNumInputs = 3;
   const intptr_t kNumTemps = 0;
@@ skipping 19 matching lines @@
     case kTypedDataInt16ArrayCid:
     case kTypedDataUint16ArrayCid:
     case kTypedDataInt32ArrayCid:
     case kTypedDataUint32ArrayCid:
       locs->set_in(2, Location::WritableRegister());
       break;
     case kTypedDataFloat32ArrayCid:
     case kTypedDataFloat64ArrayCid:  // TODO(srdjan): Support Float64 constants.
     case kTypedDataInt32x4ArrayCid:
     case kTypedDataFloat32x4ArrayCid:
+    case kTypedDataFloat64x2ArrayCid:
       locs->set_in(2, Location::RequiresFpuRegister());
       break;
     default:
       UNREACHABLE();
       return NULL;
   }
   return locs;
 }
 
 
@@ skipping 119 matching lines @@
       __ add(index.reg(), index.reg(), ShifterOperand(array));
       __ StoreSToOffset(STMP, index.reg(), 0);
       break;
     }
     case kTypedDataFloat64ArrayCid: {
       DRegister in2 = EvenDRegisterOf(locs()->in(2).fpu_reg());
       __ add(index.reg(), index.reg(), ShifterOperand(array));
       __ StoreDToOffset(in2, index.reg(), 0);
       break;
     }
+    case kTypedDataFloat64x2ArrayCid:
     case kTypedDataInt32x4ArrayCid:
     case kTypedDataFloat32x4ArrayCid: {
       QRegister in = locs()->in(2).fpu_reg();
       DRegister din0 = EvenDRegisterOf(in);
       DRegister din1 = OddDRegisterOf(in);
       __ add(index.reg(), index.reg(), ShifterOperand(array));
       __ StoreDToOffset(din0, index.reg(), 0);
       __ StoreDToOffset(din1, index.reg(), 2*kWordSize);
       break;
     }
@@ skipping 730 matching lines @@
     compiler->RestoreLiveRegisters(locs);
 
     __ b(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);
+    __ mov(locs->out().reg(), ShifterOperand(R0));
+    compiler->RestoreLiveRegisters(locs);
+
+    __ b(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 1003 matching lines @@
 
   const DRegister result_even = EvenDRegisterOf(result);
   const DRegister result_odd = OddDRegisterOf(result);
   __ LoadDFromOffset(result_even, value,
       Float32x4::value_offset() - kHeapObjectTag);
   __ LoadDFromOffset(result_odd, value,
       Float32x4::value_offset() + 2*kWordSize - kHeapObjectTag);
 }
 
 
+LocationSummary* BoxFloat64x2Instr::MakeLocationSummary(bool opt) const {
+  const intptr_t kNumInputs = 1;
+  const intptr_t kNumTemps = 1;
+  LocationSummary* summary =
+      new LocationSummary(kNumInputs,
+                          kNumTemps,
+                          LocationSummary::kCallOnSlowPath);
+  summary->set_in(0, Location::RequiresFpuRegister());
+  summary->set_temp(0, Location::RequiresRegister());
+  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();
+  QRegister value = locs()->in(0).fpu_reg();
+  DRegister value_even = EvenDRegisterOf(value);
+  DRegister value_odd = OddDRegisterOf(value);
+
+  __ TryAllocate(compiler->float64x2_class(),
+                 slow_path->entry_label(),
+                 out_reg,
+                 locs()->temp(0).reg());
+  __ Bind(slow_path->exit_label());
+
+  __ StoreDToOffset(value_even, out_reg,

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

Use one store-multiple instead of two store instruction?

+      Float32x4::value_offset() - kHeapObjectTag);

[zra, 2014/02/20 22:46:40]

Is this supposed to be Float64x2::value_offset()?

+  __ StoreDToOffset(value_odd, out_reg,
+      Float32x4::value_offset() + 2*kWordSize - kHeapObjectTag);

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

spaces around *

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

Done here and elsewhere.

+}
+
+
+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 QRegister result = locs()->out().fpu_reg();
+
+  if (value_cid != kFloat64x2Cid) {
+    const Register temp = locs()->temp(0).reg();
+    Label* deopt = compiler->AddDeoptStub(deopt_id_, kDeoptCheckClass);
+    __ tst(value, ShifterOperand(kSmiTagMask));
+    __ b(deopt, EQ);
+    __ CompareClassId(value, kFloat64x2Cid, temp);
+    __ b(deopt, NE);
+  }
+
+  const DRegister result_even = EvenDRegisterOf(result);
+  const DRegister result_odd = OddDRegisterOf(result);
+  __ LoadDFromOffset(result_even, value,

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

Load-multiple (vldmd) instead of 2 vldr? This would apply to the correctponding
Int32x4 and Float32x4 instructions as well.

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

vldmd can't be used with an offset.

+      Float64x2::value_offset() - kHeapObjectTag);
+  __ LoadDFromOffset(result_odd, value,
+      Float64x2::value_offset() + 2*kWordSize - kHeapObjectTag);
+}
+
+
 LocationSummary* BoxInt32x4Instr::MakeLocationSummary(bool opt) const {
   const intptr_t kNumInputs = 1;
   const intptr_t kNumTemps = 1;
   LocationSummary* summary =
       new LocationSummary(kNumInputs,
                           kNumTemps,
                           LocationSummary::kCallOnSlowPath);
   summary->set_in(0, Location::RequiresFpuRegister());
   summary->set_temp(0, Location::RequiresRegister());
   summary->set_out(Location::RequiresRegister());
@@ skipping 1865 matching lines @@
   compiler->GenerateCall(token_pos(),
                          &label,
                          PcDescriptors::kOther,
                          locs());
   __ Drop(2);  // Discard type arguments and receiver.
 }
 
 }  // namespace dart
 
 #endif  // defined TARGET_ARCH_ARM