X8632 Templatization completed.

src/IceTargetLoweringX8632Traits.h

Index: src/IceTargetLoweringX8632Traits.h
diff --git a/src/IceTargetLoweringX8632Traits.h b/src/IceTargetLoweringX8632Traits.h
index 4cd22fa82e3c8bdfc4c5c2774c228359fa9dc161..a775fb4ea31c4b695a90c2600ec77640e135c874 100644
--- a/src/IceTargetLoweringX8632Traits.h
+++ b/src/IceTargetLoweringX8632Traits.h
@@ -7,7 +7,7 @@
 //
 //===----------------------------------------------------------------------===//
 //
-// This file defines the X8632 Target Lowering Traits.
+// This file declares the X8632 Target Lowering Traits.
 //
 //===----------------------------------------------------------------------===//
 
@@ -19,15 +19,22 @@
 #include "IceDefs.h"
 #include "IceInst.h"
 #include "IceInstX8632.def"
+#include "IceOperand.h"
 #include "IceRegistersX8632.h"
 #include "IceTargetLoweringX8632.def"
+#include "IceTargetLowering.h"
 
 namespace Ice {
 
 class TargetX8632;
 
+namespace X8632 {
+class AssemblerX8632;
+} // end of namespace X8632
+
 namespace X86Internal {
 
+template <class Machine> struct Insts;
 template <class Machine> struct MachineTraits;
 
 template <> struct MachineTraits<TargetX8632> {
@@ -56,7 +63,7 @@ template <> struct MachineTraits<TargetX8632> {
   class Operand {
   public:
     Operand(const Operand &other)
-        : length_(other.length_), fixup_(other.fixup_) {
+        : fixup_(other.fixup_), length_(other.length_) {
       memmove(&encoding_[0], &other.encoding_[0], other.length_);
     }
 
@@ -98,7 +105,7 @@ template <> struct MachineTraits<TargetX8632> {
     AssemblerFixup *fixup() const { return fixup_; }
 
   protected:
-    Operand() : length_(0), fixup_(nullptr) {} // Needed by subclass Address.
+    Operand() : fixup_(nullptr), length_(0) {} // Needed by subclass Address.
 
     void SetModRM(int mod, GPRRegister rm) {
       assert((mod & ~3) == 0);
@@ -128,9 +135,9 @@ template <> struct MachineTraits<TargetX8632> {
     void SetFixup(AssemblerFixup *fixup) { fixup_ = fixup; }
 
   private:
-    uint8_t length_;
-    uint8_t encoding_[6];
     AssemblerFixup *fixup_;
+    uint8_t encoding_[6];
+    uint8_t length_;
 
     explicit Operand(GPRRegister reg) : fixup_(nullptr) { SetModRM(3, reg); }
 
@@ -255,6 +262,160 @@ template <> struct MachineTraits<TargetX8632> {
     End
   };
 
+  static const char *TargetName;
+
+  static IceString getRegName(SizeT RegNum, Type Ty) {
+    assert(RegNum < RegisterSet::Reg_NUM);
+    static const char *RegNames8[] = {
+#define X(val, encode, name, name16, name8, scratch, preserved, stackptr,      \
+          frameptr, isI8, isInt, isFP)                                         \
+  name8,
+        REGX8632_TABLE
+#undef X
+    };
+
+    static const char *RegNames16[] = {
+#define X(val, encode, name, name16, name8, scratch, preserved, stackptr,      \
+          frameptr, isI8, isInt, isFP)                                         \
+  name16,
+        REGX8632_TABLE
+#undef X
+    };
+
+    static const char *RegNames[] = {
+#define X(val, encode, name, name16, name8, scratch, preserved, stackptr,      \
+          frameptr, isI8, isInt, isFP)                                         \
+  name,
+        REGX8632_TABLE
+#undef X
+    };
+
+    switch (Ty) {
+    case IceType_i1:
+    case IceType_i8:
+      return RegNames8[RegNum];
+    case IceType_i16:
+      return RegNames16[RegNum];
+    default:
+      return RegNames[RegNum];
+    }
+  }
+
+  static void initRegisterSet(llvm::SmallBitVector *IntegerRegisters,
+                              llvm::SmallBitVector *IntegerRegistersI8,
+                              llvm::SmallBitVector *FloatRegisters,
+                              llvm::SmallBitVector *VectorRegisters,
+                              llvm::SmallBitVector *ScratchRegs) {
+#define X(val, encode, name, name16, name8, scratch, preserved, stackptr,      \
+          frameptr, isI8, isInt, isFP)                                         \
+  (*IntegerRegisters)[RegisterSet::val] = isInt;                               \
+  (*IntegerRegistersI8)[RegisterSet::val] = isI8;                              \
+  (*FloatRegisters)[RegisterSet::val] = isFP;                                  \
+  (*VectorRegisters)[RegisterSet::val] = isFP;                                 \
+  (*ScratchRegs)[RegisterSet::val] = scratch;
+    REGX8632_TABLE;
+#undef X
+  }
+
+  static llvm::SmallBitVector
+  getRegisterSet(TargetLowering::RegSetMask Include,
+                 TargetLowering::RegSetMask Exclude) {
+    llvm::SmallBitVector Registers(RegisterSet::Reg_NUM);
+
+#define X(val, encode, name, name16, name8, scratch, preserved, stackptr,      \
+          frameptr, isI8, isInt, isFP)                                         \
+  if (scratch && (Include & ::Ice::TargetLowering::RegSet_CallerSave))         \
+    Registers[RegisterSet::val] = true;                                        \
+  if (preserved && (Include & ::Ice::TargetLowering::RegSet_CalleeSave))       \
+    Registers[RegisterSet::val] = true;                                        \
+  if (stackptr && (Include & ::Ice::TargetLowering::RegSet_StackPointer))      \
+    Registers[RegisterSet::val] = true;                                        \
+  if (frameptr && (Include & ::Ice::TargetLowering::RegSet_FramePointer))      \
+    Registers[RegisterSet::val] = true;                                        \
+  if (scratch && (Exclude & ::Ice::TargetLowering::RegSet_CallerSave))         \
+    Registers[RegisterSet::val] = false;                                       \
+  if (preserved && (Exclude & ::Ice::TargetLowering::RegSet_CalleeSave))       \
+    Registers[RegisterSet::val] = false;                                       \
+  if (stackptr && (Exclude & ::Ice::TargetLowering::RegSet_StackPointer))      \
+    Registers[RegisterSet::val] = false;                                       \
+  if (frameptr && (Exclude & ::Ice::TargetLowering::RegSet_FramePointer))      \
+    Registers[RegisterSet::val] = false;
+
+    REGX8632_TABLE
+
+#undef X
+
+    return Registers;
+  }
+
+  static void
+  makeRandomRegisterPermutation(GlobalContext *Ctx, Cfg *Func,
+                                llvm::SmallVectorImpl<int32_t> &Permutation,
+                                const llvm::SmallBitVector &ExcludeRegisters) {
+    // TODO(stichnot): Declaring Permutation this way loses type/size
+    // information.  Fix this in conjunction with the caller-side TODO.
+    assert(Permutation.size() >= RegisterSet::Reg_NUM);
+    // Expected upper bound on the number of registers in a single equivalence
+    // class.  For x86-32, this would comprise the 8 XMM registers.  This is for
+    // performance, not correctness.
+    static const unsigned MaxEquivalenceClassSize = 8;
+    typedef llvm::SmallVector<int32_t, MaxEquivalenceClassSize> RegisterList;
+    typedef std::map<uint32_t, RegisterList> EquivalenceClassMap;
+    EquivalenceClassMap EquivalenceClasses;
+    SizeT NumShuffled = 0, NumPreserved = 0;
+
+// Build up the equivalence classes of registers by looking at the register
+// properties as well as whether the registers should be explicitly excluded
+// from shuffling.
+#define X(val, encode, name, name16, name8, scratch, preserved, stackptr,      \
+          frameptr, isI8, isInt, isFP)                                         \
+  if (ExcludeRegisters[RegisterSet::val]) {                                    \
+    /* val stays the same in the resulting permutation. */                     \
+    Permutation[RegisterSet::val] = RegisterSet::val;                          \
+    ++NumPreserved;                                                            \
+  } else {                                                                     \
+    const uint32_t Index = (scratch << 0) | (preserved << 1) | (isI8 << 2) |   \
+                           (isInt << 3) | (isFP << 4);                         \
+    /* val is assigned to an equivalence class based on its properties. */     \
+    EquivalenceClasses[Index].push_back(RegisterSet::val);                     \
+  }
+    REGX8632_TABLE
+#undef X
+
+    RandomNumberGeneratorWrapper RNG(Ctx->getRNG());
+
+    // Shuffle the resulting equivalence classes.
+    for (auto I : EquivalenceClasses) {
+      const RegisterList &List = I.second;
+      RegisterList Shuffled(List);
+      RandomShuffle(Shuffled.begin(), Shuffled.end(), RNG);
+      for (size_t SI = 0, SE = Shuffled.size(); SI < SE; ++SI) {
+        Permutation[List[SI]] = Shuffled[SI];
+        ++NumShuffled;
+      }
+    }
+
+    assert(NumShuffled + NumPreserved == RegisterSet::Reg_NUM);
+
+    if (Func->isVerbose(IceV_Random)) {
+      OstreamLocker L(Func->getContext());
+      Ostream &Str = Func->getContext()->getStrDump();
+      Str << "Register equivalence classes:\n";
+      for (auto I : EquivalenceClasses) {
+        Str << "{";
+        const RegisterList &List = I.second;
+        bool First = true;
+        for (int32_t Register : List) {
+          if (!First)
+            Str << " ";
+          First = false;
+          Str << getRegName(Register, IceType_i32);
+        }
+        Str << "}\n";
+      }
+    }
+  }
+
   // The maximum number of arguments to pass in XMM registers
   static const uint32_t X86_MAX_XMM_ARGS = 4;
   // The number of bits in a byte
@@ -287,29 +448,29 @@ template <> struct MachineTraits<TargetX8632> {
   // Note: The following data structures are defined in
   // IceTargetLoweringX8632.cpp.
 
-  // The following table summarizes the logic for lowering the fcmp
-  // instruction.  There is one table entry for each of the 16 conditions.
+  // The following table summarizes the logic for lowering the fcmp instruction.
+  // There is one table entry for each of the 16 conditions.
   //
-  // The first four columns describe the case when the operands are
-  // floating point scalar values.  A comment in lowerFcmp() describes the
-  // lowering template.  In the most general case, there is a compare
-  // followed by two conditional branches, because some fcmp conditions
-  // don't map to a single x86 conditional branch.  However, in many cases
-  // it is possible to swap the operands in the comparison and have a
-  // single conditional branch.  Since it's quite tedious to validate the
-  // table by hand, good execution tests are helpful.
+  // The first four columns describe the case when the operands are floating
+  // point scalar values.  A comment in lowerFcmp() describes the lowering
+  // template.  In the most general case, there is a compare followed by two
+  // conditional branches, because some fcmp conditions don't map to a single
+  // x86 conditional branch.  However, in many cases it is possible to swap the
+  // operands in the comparison and have a single conditional branch.  Since
+  // it's quite tedious to validate the table by hand, good execution tests are
+  // helpful.
   //
-  // The last two columns describe the case when the operands are vectors
-  // of floating point values.  For most fcmp conditions, there is a clear
-  // mapping to a single x86 cmpps instruction variant.  Some fcmp
-  // conditions require special code to handle and these are marked in the
-  // table with a Cmpps_Invalid predicate.
+  // The last two columns describe the case when the operands are vectors of
+  // floating point values.  For most fcmp conditions, there is a clear mapping
+  // to a single x86 cmpps instruction variant.  Some fcmp conditions require
+  // special code to handle and these are marked in the table with a
+  // Cmpps_Invalid predicate.
   static const struct TableFcmpType {
     uint32_t Default;
     bool SwapScalarOperands;
-    CondX86::BrCond C1, C2;
+    Cond::BrCond C1, C2;
     bool SwapVectorOperands;
-    CondX86::CmppsCond Predicate;
+    Cond::CmppsCond Predicate;
   } TableFcmp[];
   static const size_t TableFcmpSize;
 
@@ -317,9 +478,7 @@ template <> struct MachineTraits<TargetX8632> {
   // for i32 and narrower types.  Each icmp condition has a clear mapping to an
   // x86 conditional branch instruction.
 
-  static const struct TableIcmp32Type {
-    CondX86::BrCond Mapping;
-  } TableIcmp32[];
+  static const struct TableIcmp32Type { Cond::BrCond Mapping; } TableIcmp32[];
   static const size_t TableIcmp32Size;
 
   // The following table summarizes the logic for lowering the icmp instruction
@@ -327,11 +486,11 @@ template <> struct MachineTraits<TargetX8632> {
   // conditional branches are needed.  For the other conditions, three separate
   // conditional branches are needed.
   static const struct TableIcmp64Type {
-    CondX86::BrCond C1, C2, C3;
+    Cond::BrCond C1, C2, C3;
   } TableIcmp64[];
   static const size_t TableIcmp64Size;
 
-  static CondX86::BrCond getIcmp32Mapping(InstIcmp::ICond Cond) {
+  static Cond::BrCond getIcmp32Mapping(InstIcmp::ICond Cond) {
     size_t Index = static_cast<size_t>(Cond);
     assert(Index < TableIcmp32Size);
     return TableIcmp32[Index].Mapping;
@@ -341,6 +500,189 @@ template <> struct MachineTraits<TargetX8632> {
     Type InVectorElementType;
   } TableTypeX8632Attributes[];
   static const size_t TableTypeX8632AttributesSize;
+
+  //----------------------------------------------------------------------------
+  //      __  __   __  ______  ______
+  //    /\ \/\ "-.\ \/\  ___\/\__  _\
+  //    \ \ \ \ \-.  \ \___  \/_/\ \/
+  //     \ \_\ \_\\"\_\/\_____\ \ \_\
+  //      \/_/\/_/ \/_/\/_____/  \/_/
+  //
+  //----------------------------------------------------------------------------
+  using Insts = ::Ice::X86Internal::Insts<TargetX8632>;
+
+  using TargetLowering = TargetX8632;
+  using Assembler = X8632::AssemblerX8632;
+
+  // X86Operand extends the Operand hierarchy.  Its subclasses are X86OperandMem
+  // and VariableSplit.
+  class X86Operand : public ::Ice::Operand {
+    X86Operand() = delete;
+    X86Operand(const X86Operand &) = delete;
+    X86Operand &operator=(const X86Operand &) = delete;
+
+  public:
+    enum OperandKindX8632 { k__Start = ::Ice::Operand::kTarget, kMem, kSplit };
+    using ::Ice::Operand::dump;
+
+    void dump(const Cfg *, Ostream &Str) const override;
+
+  protected:
+    X86Operand(OperandKindX8632 Kind, Type Ty)
+        : Operand(static_cast<::Ice::Operand::OperandKind>(Kind), Ty) {}
+  };
+
+  // X86OperandMem represents the m32 addressing mode, with optional base and
+  // index registers, a constant offset, and a fixed shift value for the index
+  // register.
+  class X86OperandMem : public X86Operand {
+    X86OperandMem() = delete;
+    X86OperandMem(const X86OperandMem &) = delete;
+    X86OperandMem &operator=(const X86OperandMem &) = delete;
+
+  public:
+    enum SegmentRegisters {
+      DefaultSegment = -1,
+#define X(val, name, prefix) val,
+      SEG_REGX8632_TABLE
+#undef X
+          SegReg_NUM
+    };
+    static X86OperandMem *create(Cfg *Func, Type Ty, Variable *Base,
+                                 Constant *Offset, Variable *Index = nullptr,
+                                 uint16_t Shift = 0,
+                                 SegmentRegisters SegmentReg = DefaultSegment) {
+      return new (Func->allocate<X86OperandMem>())
+          X86OperandMem(Func, Ty, Base, Offset, Index, Shift, SegmentReg);
+    }
+    Variable *getBase() const { return Base; }
+    Constant *getOffset() const { return Offset; }
+    Variable *getIndex() const { return Index; }
+    uint16_t getShift() const { return Shift; }
+    SegmentRegisters getSegmentRegister() const { return SegmentReg; }
+    void emitSegmentOverride(Assembler *Asm) const;
+    Address toAsmAddress(Assembler *Asm) const;
+
+    void emit(const Cfg *Func) const override;
+    using X86Operand::dump;
+    void dump(const Cfg *Func, Ostream &Str) const override;
+
+    static bool classof(const Operand *Operand) {
+      return Operand->getKind() == static_cast<OperandKind>(kMem);
+    }
+
+    void setRandomized(bool R) { Randomized = R; }
+
+    bool getRandomized() const { return Randomized; }
+
+  private:
+    X86OperandMem(Cfg *Func, Type Ty, Variable *Base, Constant *Offset,
+                  Variable *Index, uint16_t Shift, SegmentRegisters SegmentReg);
+
+    Variable *Base;
+    Constant *Offset;
+    Variable *Index;
+    uint16_t Shift;
+    SegmentRegisters SegmentReg : 16;
+    // A flag to show if this memory operand is a randomized one. Randomized
+    // memory operands are generated in
+    // TargetX86Base::randomizeOrPoolImmediate()
+    bool Randomized;
+  };
+
+  // VariableSplit is a way to treat an f64 memory location as a pair of i32
+  // locations (Low and High).  This is needed for some cases of the Bitcast
+  // instruction.  Since it's not possible for integer registers to access the
+  // XMM registers and vice versa, the lowering forces the f64 to be spilled to
+  // the stack and then accesses through the VariableSplit.
+  // TODO(jpp): remove references to SplitVariable from IceInstX86Base as 64bit
+  // targets can natively handle these.
+  class VariableSplit : public X86Operand {
+    VariableSplit() = delete;
+    VariableSplit(const VariableSplit &) = delete;
+    VariableSplit &operator=(const VariableSplit &) = delete;
+
+  public:
+    enum Portion { Low, High };
+    static VariableSplit *create(Cfg *Func, Variable *Var, Portion Part) {
+      return new (Func->allocate<VariableSplit>())
+          VariableSplit(Func, Var, Part);
+    }
+    int32_t getOffset() const { return Part == High ? 4 : 0; }
+
+    Address toAsmAddress(const Cfg *Func) const;
+    void emit(const Cfg *Func) const override;
+    using X86Operand::dump;
+    void dump(const Cfg *Func, Ostream &Str) const override;
+
+    static bool classof(const Operand *Operand) {
+      return Operand->getKind() == static_cast<OperandKind>(kSplit);
+    }
+
+  private:
+    VariableSplit(Cfg *Func, Variable *Var, Portion Part)
+        : X86Operand(kSplit, IceType_i32), Var(Var), Part(Part) {
+      assert(Var->getType() == IceType_f64);
+      Vars = Func->allocateArrayOf<Variable *>(1);
+      Vars[0] = Var;
+      NumVars = 1;
+    }
+
+    Variable *Var;
+    Portion Part;
+  };
+
+  // SpillVariable decorates a Variable by linking it to another Variable.  When
+  // stack frame offsets are computed, the SpillVariable is given a distinct
+  // stack slot only if its linked Variable has a register.  If the linked
+  // Variable has a stack slot, then the Variable and SpillVariable share that
+  // slot.
+  class SpillVariable : public Variable {
+    SpillVariable() = delete;
+    SpillVariable(const SpillVariable &) = delete;
+    SpillVariable &operator=(const SpillVariable &) = delete;
+
+  public:
+    static SpillVariable *create(Cfg *Func, Type Ty, SizeT Index) {
+      return new (Func->allocate<SpillVariable>()) SpillVariable(Ty, Index);
+    }
+    const static OperandKind SpillVariableKind =
+        static_cast<OperandKind>(kVariable_Target);
+    static bool classof(const Operand *Operand) {
+      return Operand->getKind() == SpillVariableKind;
+    }
+    void setLinkedTo(Variable *Var) { LinkedTo = Var; }
+    Variable *getLinkedTo() const { return LinkedTo; }
+    // Inherit dump() and emit() from Variable.
+  private:
+    SpillVariable(Type Ty, SizeT Index)
+        : Variable(SpillVariableKind, Ty, Index), LinkedTo(nullptr) {}
+    Variable *LinkedTo;
+  };
+
+  // Note: The following data structures are defined in IceInstX8632.cpp.
+
+  static const struct InstBrAttributesType {
+    Cond::BrCond Opposite;
+    const char *DisplayString;
+    const char *EmitString;
+  } InstBrAttributes[];
+
+  static const struct InstCmppsAttributesType {
+    const char *EmitString;
+  } InstCmppsAttributes[];
+
+  static const struct TypeAttributesType {
+    const char *CvtString;   // i (integer), s (single FP), d (double FP)
+    const char *SdSsString;  // ss, sd, or <blank>
+    const char *PackString;  // b, w, d, or <blank>
+    const char *WidthString; // b, w, l, q, or <blank>
+    const char *FldString;   // s, l, or <blank>
+  } TypeAttributes[];
+
+  static const char *InstSegmentRegNames[];
+
+  static uint8_t InstSegmentPrefixes[];
 };
 
 } // end of namespace X86Internal