[turbofan] Implemented the TruncateFloat32ToInt64 TurboFan operator.

The TruncateFloat32ToInt64 operator converts a float32 to an int64 using
the round-to-zero rounding mode (truncate). If the input value is
outside the int64 range, then the result depends on the architecture. I
implemented the operator on x64, arm64, and mips64.

R=titzer@chromium.org, jacob.bramley@arm.com

Committed: https://crrev.com/1df1066c3c77464d2a68d7c8d501a5a0f3ad195a
Cr-Commit-Position: refs/heads/master@{#32315}

Review URL: https://codereview.chromium.org/1476063002

Cr-Commit-Position: refs/heads/master@{#32325}

src/compiler/arm64/code-generator-arm64.cc

@@ -1022,6 +1022,9 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
     case kArm64Float64ToUint32:
       __ Fcvtzu(i.OutputRegister32(), i.InputDoubleRegister(0));
       break;
+    case kArm64Float32ToInt64:
+      __ Fcvtzs(i.OutputRegister64(), i.InputFloat32Register(0));
+      break;
     case kArm64Float64ToInt64:
       __ Fcvtzs(i.OutputRegister64(), i.InputDoubleRegister(0));
       break;

src/compiler/arm64/instruction-codes-arm64.h

@@ -112,6 +112,7 @@ namespace compiler {
   V(Arm64Float64ToFloat32)         \
   V(Arm64Float64ToInt32)           \
   V(Arm64Float64ToUint32)          \
+  V(Arm64Float32ToInt64)           \
   V(Arm64Float64ToInt64)           \
   V(Arm64Float64ToUint64)          \
   V(Arm64Int32ToFloat64)           \

src/compiler/arm64/instruction-selector-arm64.cc

@@ -1237,6 +1237,11 @@ void InstructionSelector::VisitChangeFloat64ToUint32(Node* node) {
 }
 
 
+void InstructionSelector::VisitTruncateFloat32ToInt64(Node* node) {
+  VisitRR(this, kArm64Float32ToInt64, node);
+}
+
+
 void InstructionSelector::VisitTruncateFloat64ToInt64(Node* node) {
   VisitRR(this, kArm64Float64ToInt64, node);
 }

src/compiler/instruction-selector.cc

@@ -847,6 +847,8 @@ void InstructionSelector::VisitNode(Node* node) {
       return MarkAsWord32(node), VisitChangeFloat64ToInt32(node);
     case IrOpcode::kChangeFloat64ToUint32:
       return MarkAsWord32(node), VisitChangeFloat64ToUint32(node);
+    case IrOpcode::kTruncateFloat32ToInt64:
+      return MarkAsWord64(node), VisitTruncateFloat32ToInt64(node);
     case IrOpcode::kTruncateFloat64ToInt64:
       return MarkAsWord64(node), VisitTruncateFloat64ToInt64(node);
     case IrOpcode::kTruncateFloat64ToUint64:
@@ -1099,6 +1101,11 @@ void InstructionSelector::VisitChangeUint32ToUint64(Node* node) {
 }
 
 
+void InstructionSelector::VisitTruncateFloat32ToInt64(Node* node) {
+  UNIMPLEMENTED();
+}
+
+
 void InstructionSelector::VisitTruncateFloat64ToInt64(Node* node) {
   UNIMPLEMENTED();
 }

src/compiler/machine-operator.cc

@@ -135,6 +135,7 @@ CheckedStoreRepresentation CheckedStoreRepresentationOf(Operator const* op) {
   V(ChangeFloat32ToFloat64, Operator::kNoProperties, 1, 0, 1)                 \
   V(ChangeFloat64ToInt32, Operator::kNoProperties, 1, 0, 1)                   \
   V(ChangeFloat64ToUint32, Operator::kNoProperties, 1, 0, 1)                  \
+  V(TruncateFloat32ToInt64, Operator::kNoProperties, 1, 0, 1)                 \
   V(TruncateFloat64ToInt64, Operator::kNoProperties, 1, 0, 1)                 \
   V(TruncateFloat64ToUint64, Operator::kNoProperties, 1, 0, 1)                \
   V(ChangeInt32ToFloat64, Operator::kNoProperties, 1, 0, 1)                   \

src/compiler/machine-operator.h

@@ -207,6 +207,7 @@ class MachineOperatorBuilder final : public ZoneObject {
   const Operator* ChangeFloat32ToFloat64();
   const Operator* ChangeFloat64ToInt32();   // narrowing
   const Operator* ChangeFloat64ToUint32();  // narrowing
+  const Operator* TruncateFloat32ToInt64();
   const Operator* TruncateFloat64ToInt64();
   const Operator* TruncateFloat64ToUint64();
   const Operator* ChangeInt32ToFloat64();

src/compiler/mips64/code-generator-mips64.cc

@@ -991,6 +991,13 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       __ mfc1(i.OutputRegister(), scratch);
       break;
     }
+    case kMips64TruncLS: {
+      FPURegister scratch = kScratchDoubleReg;
+      // Other arches use round to zero here, so we follow.
+      __ trunc_l_s(scratch, i.InputDoubleRegister(0));
+      __ dmfc1(i.OutputRegister(), scratch);
+      break;
+    }
     case kMips64TruncLD: {
       FPURegister scratch = kScratchDoubleReg;
       // Other arches use round to zero here, so we follow.

src/compiler/mips64/instruction-codes-mips64.h

@@ -75,6 +75,7 @@ namespace compiler {
   V(Mips64CvtSD)                    \
   V(Mips64CvtDS)                    \
   V(Mips64TruncWD)                  \
+  V(Mips64TruncLS)                  \
   V(Mips64TruncLD)                  \
   V(Mips64TruncUwD)                 \
   V(Mips64TruncUlD)                 \

src/compiler/mips64/instruction-selector-mips64.cc

@@ -731,6 +731,11 @@ void InstructionSelector::VisitChangeFloat64ToUint32(Node* node) {
 }
 
 
+void InstructionSelector::VisitTruncateFloat32ToInt64(Node* node) {
+  VisitRR(this, kMips64TruncLS, node);
+}
+
+
 void InstructionSelector::VisitTruncateFloat64ToInt64(Node* node) {
   VisitRR(this, kMips64TruncLD, node);
 }

src/compiler/opcodes.h

@@ -270,6 +270,7 @@
   V(ChangeFloat32ToFloat64)     \
   V(ChangeFloat64ToInt32)       \
   V(ChangeFloat64ToUint32)      \
+  V(TruncateFloat32ToInt64)     \
   V(TruncateFloat64ToInt64)     \
   V(TruncateFloat64ToUint64)    \
   V(ChangeInt32ToFloat64)       \

src/compiler/operator.h

@@ -30,7 +30,7 @@ namespace compiler {
 // meaningful to the operator itself.
 class Operator : public ZoneObject {
  public:
-  typedef uint8_t Opcode;
+  typedef uint16_t Opcode;
 
   // Properties inform the operator-independent optimizer about legal
   // transformations for nodes that have this operator.

src/compiler/ppc/instruction-selector-ppc.cc

@@ -927,6 +927,11 @@ void InstructionSelector::VisitChangeFloat64ToUint32(Node* node) {
 
 
 #if V8_TARGET_ARCH_PPC64
+void InstructionSelector::VisitTruncateFloat32ToInt64(Node* node) {
+  UNIMPLEMENTED();
+}
+
+
 void InstructionSelector::VisitTruncateFloat64ToInt64(Node* node) {
   VisitRR(this, kPPC_DoubleToInt64, node);
 }

src/compiler/raw-machine-assembler.h

@@ -438,6 +438,9 @@ class RawMachineAssembler {
   Node* ChangeFloat64ToUint32(Node* a) {
     return AddNode(machine()->ChangeFloat64ToUint32(), a);
   }
+  Node* TruncateFloat32ToInt64(Node* a) {
+    return AddNode(machine()->TruncateFloat32ToInt64(), a);
+  }
   Node* TruncateFloat64ToInt64(Node* a) {
     return AddNode(machine()->TruncateFloat64ToInt64(), a);
   }

src/compiler/typer.cc

@@ -2102,6 +2102,11 @@ Type* Typer::Visitor::TypeChangeFloat64ToUint32(Node* node) {
 }
 
 
+Type* Typer::Visitor::TypeTruncateFloat32ToInt64(Node* node) {
+  return Type::Internal();
+}
+
+
 Type* Typer::Visitor::TypeTruncateFloat64ToInt64(Node* node) {
   return Type::Internal();
 }

src/compiler/verifier.cc

@@ -919,6 +919,7 @@ void Verifier::Visitor::Check(Node* node) {
     case IrOpcode::kChangeFloat32ToFloat64:
     case IrOpcode::kChangeFloat64ToInt32:
     case IrOpcode::kChangeFloat64ToUint32:
+    case IrOpcode::kTruncateFloat32ToInt64:
     case IrOpcode::kTruncateFloat64ToInt64:
     case IrOpcode::kTruncateFloat64ToUint64:
     case IrOpcode::kFloat64ExtractLowWord32:

src/compiler/x64/code-generator-x64.cc

@@ -1043,6 +1043,13 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       __ AssertZeroExtended(i.OutputRegister());
       break;
     }
+    case kSSEFloat32ToInt64:
+      if (instr->InputAt(0)->IsDoubleRegister()) {
+        __ Cvttss2siq(i.OutputRegister(), i.InputDoubleRegister(0));
+      } else {
+        __ Cvttss2siq(i.OutputRegister(), i.InputOperand(0));
+      }
+      break;
     case kSSEFloat64ToInt64:
       if (instr->InputAt(0)->IsDoubleRegister()) {
         __ Cvttsd2siq(i.OutputRegister(), i.InputDoubleRegister(0));

src/compiler/x64/instruction-codes-x64.h

@@ -79,6 +79,7 @@ namespace compiler {
   V(SSEFloat64ToFloat32)           \
   V(SSEFloat64ToInt32)             \
   V(SSEFloat64ToUint32)            \
+  V(SSEFloat32ToInt64)             \
   V(SSEFloat64ToInt64)             \
   V(SSEFloat64ToUint64)            \
   V(SSEInt32ToFloat64)             \

src/compiler/x64/instruction-selector-x64.cc

@@ -834,6 +834,12 @@ void InstructionSelector::VisitChangeFloat64ToUint32(Node* node) {
 }
 
 
+void InstructionSelector::VisitTruncateFloat32ToInt64(Node* node) {
+  X64OperandGenerator g(this);
+  Emit(kSSEFloat32ToInt64, g.DefineAsRegister(node), g.Use(node->InputAt(0)));
+}
+
+
 void InstructionSelector::VisitTruncateFloat64ToInt64(Node* node) {
   X64OperandGenerator g(this);
   Emit(kSSEFloat64ToInt64, g.DefineAsRegister(node), g.Use(node->InputAt(0)));

src/x64/assembler-x64.cc

@@ -3112,6 +3112,28 @@ void Assembler::cvttsd2si(Register dst, XMMRegister src) {
 }
 
 
+void Assembler::cvttss2siq(Register dst, XMMRegister src) {
+  DCHECK(!IsEnabled(AVX));
+  EnsureSpace ensure_space(this);
+  emit(0xF3);
+  emit_rex_64(dst, src);
+  emit(0x0F);
+  emit(0x2C);
+  emit_sse_operand(dst, src);
+}
+
+
+void Assembler::cvttss2siq(Register dst, const Operand& src) {
+  DCHECK(!IsEnabled(AVX));
+  EnsureSpace ensure_space(this);
+  emit(0xF3);
+  emit_rex_64(dst, src);
+  emit(0x0F);
+  emit(0x2C);
+  emit_sse_operand(dst, src);
+}
+
+
 void Assembler::cvttsd2siq(Register dst, XMMRegister src) {
   DCHECK(!IsEnabled(AVX));
   EnsureSpace ensure_space(this);

src/x64/assembler-x64.h

@@ -1077,6 +1077,8 @@ class Assembler : public AssemblerBase {
 
   void cvttsd2si(Register dst, const Operand& src);
   void cvttsd2si(Register dst, XMMRegister src);
+  void cvttss2siq(Register dst, XMMRegister src);
+  void cvttss2siq(Register dst, const Operand& src);
   void cvttsd2siq(Register dst, XMMRegister src);
   void cvttsd2siq(Register dst, const Operand& src);
 
@@ -1390,6 +1392,14 @@ class Assembler : public AssemblerBase {
     XMMRegister idst = {dst.code()};
     vsd(0x2c, idst, xmm0, src, kF2, k0F, kW0);
   }
+  void vcvttss2siq(Register dst, XMMRegister src) {
+    XMMRegister idst = {dst.code()};
+    vsd(0x2c, idst, xmm0, src, kF3, k0F, kW1);
+  }
+  void vcvttss2siq(Register dst, const Operand& src) {
+    XMMRegister idst = {dst.code()};
+    vsd(0x2c, idst, xmm0, src, kF3, k0F, kW1);
+  }
   void vcvttsd2siq(Register dst, XMMRegister src) {
     XMMRegister idst = {dst.code()};
     vsd(0x2c, idst, xmm0, src, kF2, k0F, kW1);

src/x64/disasm-x64.cc

@@ -989,6 +989,11 @@ int DisassemblerX64::AVXInstruction(byte* data) {
                        NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
         current += PrintRightOperand(current);
         break;
+      case 0x2c:
+        AppendToBuffer("vcvttss2si%s %s,", vex_w() ? "q" : "",
+                       NameOfCPURegister(regop));
+        current += PrintRightXMMOperand(current);
+        break;
       case 0x58:
         AppendToBuffer("vaddss %s,%s,", NameOfXMMRegister(regop),
                        NameOfXMMRegister(vvvv));

src/x64/macro-assembler-x64.cc

@@ -967,6 +967,26 @@ void MacroAssembler::Cvttsd2si(Register dst, const Operand& src) {
 }
 
 
+void MacroAssembler::Cvttss2siq(Register dst, XMMRegister src) {
+  if (CpuFeatures::IsSupported(AVX)) {
+    CpuFeatureScope scope(this, AVX);
+    vcvttss2siq(dst, src);
+  } else {
+    cvttss2siq(dst, src);
+  }
+}
+
+
+void MacroAssembler::Cvttss2siq(Register dst, const Operand& src) {
+  if (CpuFeatures::IsSupported(AVX)) {
+    CpuFeatureScope scope(this, AVX);
+    vcvttss2siq(dst, src);
+  } else {
+    cvttss2siq(dst, src);
+  }
+}
+
+
 void MacroAssembler::Cvttsd2siq(Register dst, XMMRegister src) {
   if (CpuFeatures::IsSupported(AVX)) {
     CpuFeatureScope scope(this, AVX);

src/x64/macro-assembler-x64.h

@@ -832,6 +832,8 @@ class MacroAssembler: public Assembler {
 
   void Cvttsd2si(Register dst, XMMRegister src);
   void Cvttsd2si(Register dst, const Operand& src);
+  void Cvttss2siq(Register dst, XMMRegister src);
+  void Cvttss2siq(Register dst, const Operand& src);
   void Cvttsd2siq(Register dst, XMMRegister src);
   void Cvttsd2siq(Register dst, const Operand& src);
 

test/cctest/compiler/test-run-machops.cc

@@ -5391,6 +5391,24 @@ TEST(RunBitcastFloat64ToInt64) {
 }
 
 
+TEST(RunTruncateFloat32ToInt64) {
+  BufferedRawMachineAssemblerTester<int64_t> m(kMachFloat32);
+  m.Return(m.TruncateFloat32ToInt64(m.Parameter(0)));
+
+  FOR_INT64_INPUTS(i) {
+    float input = static_cast<float>(*i);
+    if (input < 9223372036854775808.0 && input > -9223372036854775809.0) {
+      CHECK_EQ(static_cast<int64_t>(input), m.Call(input));
+    }
+  }
+  FOR_FLOAT32_INPUTS(j) {
+    if (*j < 9223372036854775808.0 && *j > -9223372036854775809.0) {
+      CHECK_EQ(static_cast<int64_t>(*j), m.Call(*j));
+    }
+  }
+}
+
+
 TEST(RunTruncateFloat64ToInt64) {
   BufferedRawMachineAssemblerTester<int64_t> m(kMachFloat64);
   m.Return(m.TruncateFloat64ToInt64(m.Parameter(0)));