Explicit conversions for Float32 array loads/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 884 matching lines @@
   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) ||
       (representation() == kUnboxedFloat32x4) ||
       (representation() == kUnboxedInt32x4)) {
-    locs->set_out(Location::RequiresFpuRegister());
+    if (class_id() == kTypedDataFloat32ArrayCid) {
+      // Need register <= Q3 for float operations.
+      locs->set_out(Location::FpuRegisterLocation(Q3));
+    } else {
+      locs->set_out(Location::RequiresFpuRegister());
+    }
   } else {
     locs->set_out(Location::RequiresRegister());
   }
   return locs;
 }
 
 
 void LoadIndexedInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
   Register array = locs()->in(0).reg();
   Location index = locs()->in(1);
@@ skipping 43 matching lines @@
     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.
+        // Load single precision float.
         // 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);
+        __ vldrs(EvenSRegisterOf(dresult0), element_address);
         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 kTypedDataInt32x4ArrayCid:
       case kTypedDataFloat32x4ArrayCid:
         __ add(index.reg(), index.reg(), ShifterOperand(array));
@@ skipping 111 matching lines @@
     case kTypedDataUint8ArrayCid:
     case kTypedDataUint8ClampedArrayCid:
     case kOneByteStringCid:
     case kTypedDataInt16ArrayCid:
     case kTypedDataUint16ArrayCid:
     case kTypedDataInt32ArrayCid:
     case kTypedDataUint32ArrayCid:
       locs->set_in(2, Location::WritableRegister());
       break;
     case kTypedDataFloat32ArrayCid:
+      // Need low register (<= Q3).
+      locs->set_in(2, Location::FpuRegisterLocation(Q3));
+      break;
     case kTypedDataFloat64ArrayCid:  // TODO(srdjan): Support Float64 constants.
     case kTypedDataInt32x4ArrayCid:
     case kTypedDataFloat32x4ArrayCid:
       locs->set_in(2, Location::RequiresFpuRegister());
       break;
     default:
       UNREACHABLE();
       return NULL;
   }
   return locs;
@@ skipping 108 matching lines @@
         ASSERT(RequiredInputRepresentation(2) == kTagged);
         Register value = locs()->in(2).reg();
         __ SmiUntag(value);
         __ str(value, element_address);
       } else {
         UNIMPLEMENTED();
       }
       break;
     }
     case kTypedDataFloat32ArrayCid: {
-      DRegister in2 = EvenDRegisterOf(locs()->in(2).fpu_reg());
-      // Convert to single precision.
-      __ vcvtsd(STMP, in2);
-      // Store.
+      SRegister value =
+          EvenSRegisterOf(EvenDRegisterOf(locs()->in(2).fpu_reg()));
       __ add(index.reg(), index.reg(), ShifterOperand(array));
-      __ StoreSToOffset(STMP, index.reg(), 0);
+      __ StoreSToOffset(value, 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 kTypedDataInt32x4ArrayCid:
     case kTypedDataFloat32x4ArrayCid: {
@@ skipping 2953 matching lines @@
   __ vcvtid(STMP, value);
   __ vmovrs(result, STMP);
   // Check for overflow and that it fits into Smi.
   __ CompareImmediate(result, 0xC0000000);
   __ b(deopt, MI);
   __ SmiTag(result);
 }
 
 
 LocationSummary* DoubleToDoubleInstr::MakeLocationSummary(bool opt) const {
+  UNIMPLEMENTED();
+  return NULL;
+}
+
+
+void DoubleToDoubleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  UNIMPLEMENTED();
+}
+
+
+LocationSummary* DoubleToFloatInstr::MakeLocationSummary(bool opt) const {
   const intptr_t kNumInputs = 1;
   const intptr_t kNumTemps = 0;
   LocationSummary* result =
       new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
-  result->set_in(0, Location::RequiresFpuRegister());
-  result->set_out(Location::RequiresFpuRegister());
+  // Low (<= Q3) Q registers are needed for the conversion instructions.
+  result->set_in(0, Location::FpuRegisterLocation(Q3));
+  result->set_out(Location::SameAsFirstInput());
   return result;
 }
 
 
-void DoubleToDoubleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
-  // QRegister value = locs()->in(0).fpu_reg();
-  // QRegister result = locs()->out().fpu_reg();
-  switch (recognized_kind()) {
-    case MethodRecognizer::kDoubleTruncate:
-      UNIMPLEMENTED();
-      // __ roundsd(result, value,  Assembler::kRoundToZero);
-      break;
-    case MethodRecognizer::kDoubleFloor:
-      UNIMPLEMENTED();
-      // __ roundsd(result, value,  Assembler::kRoundDown);
-      break;
-    case MethodRecognizer::kDoubleCeil:
-      UNIMPLEMENTED();
-      // __ roundsd(result, value,  Assembler::kRoundUp);
-      break;
-    default:
-      UNREACHABLE();
-  }
+void DoubleToFloatInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  DRegister value = EvenDRegisterOf(locs()->in(0).fpu_reg());
+  SRegister result = EvenSRegisterOf(EvenDRegisterOf(locs()->out().fpu_reg()));
+  __ vcvtsd(result, value);
+}
+
+
+LocationSummary* FloatToDoubleInstr::MakeLocationSummary(bool opt) const {
+  const intptr_t kNumInputs = 1;
+  const intptr_t kNumTemps = 0;
+  LocationSummary* result =
+      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
+  // Low (<= Q3) Q registers are needed for the conversion instructions.
+  result->set_in(0, Location::FpuRegisterLocation(Q3));
+  result->set_out(Location::SameAsFirstInput());
+  return result;
+}
+
+
+void FloatToDoubleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  SRegister value = EvenSRegisterOf(EvenDRegisterOf(locs()->in(0).fpu_reg()));
+  DRegister result = EvenDRegisterOf(locs()->out().fpu_reg());
+  __ vcvtds(result, value);
 }
 
 
 LocationSummary* InvokeMathCFunctionInstr::MakeLocationSummary(bool opt) const {
   ASSERT((InputCount() == 1) || (InputCount() == 2));
   const intptr_t kNumTemps = 0;
   LocationSummary* result =
       new LocationSummary(InputCount(), kNumTemps, LocationSummary::kCall);
   result->set_in(0, Location::FpuRegisterLocation(Q0));
   if (InputCount() == 2) {
@@ skipping 644 matching lines @@
   compiler->GenerateCall(token_pos(),
                          &label,
                          PcDescriptors::kOther,
                          locs());
   __ Drop(2);  // Discard type arguments and receiver.
 }
 
 }  // namespace dart
 
 #endif  // defined TARGET_ARCH_ARM