Subzero: Fix some issues related to legalization and undef handling.

src/IceTargetLoweringX8632.cpp

 //===- subzero/src/IceTargetLoweringX8632.cpp - x86-32 lowering -----------===//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This file implements the TargetLoweringX8632 class, which
@@ skipping 1231 matching lines @@
       _movp(Dest, T);
     } break;
     case InstArithmetic::Mul: {
       bool TypesAreValidForPmull =
           Dest->getType() == IceType_v4i32 || Dest->getType() == IceType_v8i16;
       bool InstructionSetIsValidForPmull =
           Dest->getType() == IceType_v8i16 || InstructionSet >= SSE4_1;
       if (TypesAreValidForPmull && InstructionSetIsValidForPmull) {
         Variable *T = makeReg(Dest->getType());
         _movp(T, Src0);
-        _pmull(T, legalizeToVar(Src1));
+        _pmull(T, LEGAL_HACK(Src1));
         _movp(Dest, T);
       } else if (Dest->getType() == IceType_v4i32) {
         // Lowering sequence:
         // Note: The mask arguments have index 0 on the left.
         //
         // movups  T1, Src0
         // pshufd  T2, Src0, {1,0,3,0}
         // pshufd  T3, Src1, {1,0,3,0}
         // # T1 = {Src0[0] * Src1[0], Src0[2] * Src1[2]}
         // pmuludq T1, Src1
@@ skipping 889 matching lines @@
     case IceType_v4f32: {
       _movp(Dest, legalizeToVar(Src0));
     } break;
     }
     break;
   }
   }
 }
 
 void TargetX8632::lowerExtractElement(const InstExtractElement *Inst) {
-  Operand *SourceVectOperand = Inst->getSrc(0);
+  Operand *SourceVectNotLegalized = Inst->getSrc(0);
   ConstantInteger *ElementIndex =
       llvm::dyn_cast<ConstantInteger>(Inst->getSrc(1));
   // Only constant indices are allowed in PNaCl IR.
   assert(ElementIndex);
 
   unsigned Index = ElementIndex->getValue();
-  Type Ty = SourceVectOperand->getType();
+  Type Ty = SourceVectNotLegalized->getType();
   Type ElementTy = typeElementType(Ty);
   Type InVectorElementTy = getInVectorElementType(Ty);
-  Variable *ExtractedElement = makeReg(InVectorElementTy);
+  Variable *ExtractedElementR = makeReg(InVectorElementTy);
 
   // TODO(wala): Determine the best lowering sequences for each type.
   bool CanUsePextr =
       Ty == IceType_v8i16 || Ty == IceType_v8i1 || InstructionSet >= SSE4_1;
   if (CanUsePextr && Ty != IceType_v4f32) {
     // Use pextrb, pextrw, or pextrd.
     Constant *Mask = Ctx->getConstantInt(IceType_i8, Index);
-    Variable *SourceVectR = legalizeToVar(SourceVectOperand);
-    _pextr(ExtractedElement, SourceVectR, Mask);
+    Variable *SourceVectR = legalizeToVar(SourceVectNotLegalized);
+    _pextr(ExtractedElementR, SourceVectR, Mask);
   } else if (Ty == IceType_v4i32 || Ty == IceType_v4f32 || Ty == IceType_v4i1) {
     // Use pshufd and movd/movss.
     //
     // ALIGNHACK: Force vector operands to registers in instructions that
     // require aligned memory operands until support for stack alignment
     // is implemented.
 #define ALIGN_HACK(Vect) legalizeToVar((Vect))
+    Operand *SourceVectRM =
+        legalize(SourceVectNotLegalized, Legal_Reg | Legal_Mem);
     Variable *T = NULL;
     if (Index) {
       // The shuffle only needs to occur if the element to be extracted
       // is not at the lowest index.
       Constant *Mask = Ctx->getConstantInt(IceType_i8, Index);
       T = makeReg(Ty);
-      _pshufd(T, ALIGN_HACK(SourceVectOperand), Mask);
+      _pshufd(T, ALIGN_HACK(SourceVectRM), Mask);
     } else {
-      T = legalizeToVar(SourceVectOperand);
+      T = ALIGN_HACK(SourceVectRM);
     }
 
     if (InVectorElementTy == IceType_i32) {
-      _movd(ExtractedElement, T);
+      _movd(ExtractedElementR, T);
     } else { // Ty == Icetype_f32
       // TODO(wala): _movss is only used here because _mov does not
       // allow a vector source and a scalar destination.  _mov should be
       // able to be used here.
       // _movss is a binary instruction, so the FakeDef is needed to
       // keep the live range analysis consistent.
-      Context.insert(InstFakeDef::create(Func, ExtractedElement));
-      _movss(ExtractedElement, T);
+      Context.insert(InstFakeDef::create(Func, ExtractedElementR));
+      _movss(ExtractedElementR, T);
     }
 #undef ALIGN_HACK
   } else {
     assert(Ty == IceType_v16i8 || Ty == IceType_v16i1);
     // Spill the value to a stack slot and do the extraction in memory.
     //
-    // TODO(wala): use legalize(SourceVectOperand, Legal_Mem) when
+    // TODO(wala): use legalize(SourceVectNotLegalized, Legal_Mem) when
     // support for legalizing to mem is implemented.
     Variable *Slot = Func->makeVariable(Ty, Context.getNode());
     Slot->setWeight(RegWeight::Zero);
-    _movp(Slot, legalizeToVar(SourceVectOperand));
+    _movp(Slot, legalizeToVar(SourceVectNotLegalized));
 
     // Compute the location of the element in memory.
     unsigned Offset = Index * typeWidthInBytes(InVectorElementTy);
     OperandX8632Mem *Loc =
         getMemoryOperandForStackSlot(InVectorElementTy, Slot, Offset);
-    _mov(ExtractedElement, Loc);
+    _mov(ExtractedElementR, Loc);
   }
 
   if (ElementTy == IceType_i1) {
     // Truncate extracted integers to i1s if necessary.
     Variable *T = makeReg(IceType_i1);
     InstCast *Cast =
-        InstCast::create(Func, InstCast::Trunc, T, ExtractedElement);
+        InstCast::create(Func, InstCast::Trunc, T, ExtractedElementR);
     lowerCast(Cast);
-    ExtractedElement = T;
+    ExtractedElementR = T;
   }
 
   // Copy the element to the destination.
   Variable *Dest = Inst->getDest();
-  _mov(Dest, ExtractedElement);
+  _mov(Dest, ExtractedElementR);
 }
 
 void TargetX8632::lowerFcmp(const InstFcmp *Inst) {
   Operand *Src0 = Inst->getSrc(0);
   Operand *Src1 = Inst->getSrc(1);
   Variable *Dest = Inst->getDest();
 
   if (isVectorType(Dest->getType())) {
     InstFcmp::FCond Condition = Inst->getCondition();
     size_t Index = static_cast<size_t>(Condition);
     assert(Index < TableFcmpSize);
 
     if (TableFcmp[Index].SwapVectorOperands) {
       Operand *T = Src0;
       Src0 = Src1;
       Src1 = T;
     }
 
     Variable *T = NULL;
 
-    // ALIGNHACK: Without support for stack alignment, both operands to
-    // cmpps need to be forced into registers.  Once support for stack
-    // alignment is implemented, remove LEGAL_HACK.
+    if (Condition == InstFcmp::True) {
+      // makeVectorOfOnes() requires an integer vector type.
+      T = makeVectorOfMinusOnes(IceType_v4i32);
+    } else if (Condition == InstFcmp::False) {
+      T = makeVectorOfZeros(Dest->getType());
+    } else {
+      Operand *Src0RM = legalizeToVar(Src0, Legal_Reg | Legal_Mem);
+      Operand *Src1RM = legalizeToVar(Src1, Legal_Reg | Legal_Mem);
+
+      // ALIGNHACK: Without support for stack alignment, both operands to
+      // cmpps need to be forced into registers.  Once support for stack
+      // alignment is implemented, remove LEGAL_HACK.
 #define LEGAL_HACK(Vect) legalizeToVar((Vect))
-    switch (Condition) {
-    default: {
-      InstX8632Cmpps::CmppsCond Predicate = TableFcmp[Index].Predicate;
-      assert(Predicate != InstX8632Cmpps::Cmpps_Invalid);
-      T = makeReg(Src0->getType());
-      _movp(T, Src0);
-      _cmpps(T, LEGAL_HACK(Src1), Predicate);
-    } break;
-    case InstFcmp::False:
-      T = makeVectorOfZeros(Src0->getType());
-      break;
-    case InstFcmp::One: {
-      // Check both unequal and ordered.
-      T = makeReg(Src0->getType());
-      Variable *T2 = makeReg(Src0->getType());
-      Src1 = LEGAL_HACK(Src1);
-      _movp(T, Src0);
-      _cmpps(T, Src1, InstX8632Cmpps::Cmpps_neq);
-      _movp(T2, Src0);
-      _cmpps(T2, Src1, InstX8632Cmpps::Cmpps_ord);
-      _pand(T, T2);
-    } break;
-    case InstFcmp::Ueq: {
-      // Check both equal or unordered.
-      T = makeReg(Src0->getType());
-      Variable *T2 = makeReg(Src0->getType());
-      Src1 = LEGAL_HACK(Src1);
-      _movp(T, Src0);
-      _cmpps(T, Src1, InstX8632Cmpps::Cmpps_eq);
-      _movp(T2, Src0);
-      _cmpps(T2, Src1, InstX8632Cmpps::Cmpps_unord);
-      _por(T, T2);
-    } break;
-    case InstFcmp::True:
-      T = makeVectorOfMinusOnes(IceType_v4i32);
-      break;
+      switch (Condition) {
+      default: {
+        InstX8632Cmpps::CmppsCond Predicate = TableFcmp[Index].Predicate;
+        assert(Predicate != InstX8632Cmpps::Cmpps_Invalid);
+        T = makeReg(Src0RM->getType());
+        _movp(T, Src0RM);
+        _cmpps(T, LEGAL_HACK(Src1RM), Predicate);
+      } break;
+      case InstFcmp::One: {
+        // Check both unequal and ordered.
+        T = makeReg(Src0RM->getType());
+        Variable *T2 = makeReg(Src0RM->getType());
+        Src1RM = LEGAL_HACK(Src1RM);
+        _movp(T, Src0RM);
+        _cmpps(T, Src1RM, InstX8632Cmpps::Cmpps_neq);
+        _movp(T2, Src0RM);
+        _cmpps(T2, Src1RM, InstX8632Cmpps::Cmpps_ord);
+        _pand(T, T2);
+      } break;
+      case InstFcmp::Ueq: {
+        // Check both equal or unordered.
+        T = makeReg(Src0RM->getType());
+        Variable *T2 = makeReg(Src0RM->getType());
+        Src1RM = LEGAL_HACK(Src1RM);
+        _movp(T, Src0RM);
+        _cmpps(T, Src1RM, InstX8632Cmpps::Cmpps_eq);
+        _movp(T2, Src0RM);
+        _cmpps(T2, Src1RM, InstX8632Cmpps::Cmpps_unord);
+        _por(T, T2);
+      } break;
+      }
+#undef LEGAL_HACK
     }
-#undef LEGAL_HACK
 
     _movp(Dest, T);
     eliminateNextVectorSextInstruction(Dest);
     return;
   }
 
   // Lowering a = fcmp cond, b, c
   //   ucomiss b, c       /* only if C1 != Br_None */
   //                      /* but swap b,c order if SwapOperands==true */
   //   mov a, <default>
@@ skipping 64 matching lines @@
       Variable *NewSrc1 = Func->makeVariable(NewTy, Context.getNode());
       lowerCast(InstCast::create(Func, InstCast::Sext, NewSrc0, Src0));
       lowerCast(InstCast::create(Func, InstCast::Sext, NewSrc1, Src1));
       Src0 = NewSrc0;
       Src1 = NewSrc1;
       Ty = NewTy;
     }
 
     InstIcmp::ICond Condition = Inst->getCondition();
 
+    Operand *Src0RM = legalize(Src0, Legal_Reg | Legal_Mem);
+    Operand *Src1RM = legalize(Src1, Legal_Reg | Legal_Mem);
+
     // SSE2 only has signed comparison operations.  Transform unsigned
     // inputs in a manner that allows for the use of signed comparison
     // operations by flipping the high order bits.
     if (Condition == InstIcmp::Ugt || Condition == InstIcmp::Uge ||
         Condition == InstIcmp::Ult || Condition == InstIcmp::Ule) {
       Variable *T0 = makeReg(Ty);
       Variable *T1 = makeReg(Ty);
       Variable *HighOrderBits = makeVectorOfHighOrderBits(Ty);
-      _movp(T0, Src0);
+      _movp(T0, Src0RM);
       _pxor(T0, HighOrderBits);
-      _movp(T1, Src1);
+      _movp(T1, Src1RM);
       _pxor(T1, HighOrderBits);
-      Src0 = T0;
-      Src1 = T1;
+      Src0RM = T0;
+      Src1RM = T1;
     }
 
     // TODO: ALIGNHACK: Both operands to compare instructions need to be
     // in registers until stack alignment support is implemented.  Once
     // there is support for stack alignment, LEGAL_HACK can be removed.
 #define LEGAL_HACK(Vect) legalizeToVar((Vect))
     Variable *T = makeReg(Ty);
     switch (Condition) {
     default:
       llvm_unreachable("unexpected condition");
       break;
     case InstIcmp::Eq: {
-      _movp(T, Src0);
-      _pcmpeq(T, LEGAL_HACK(Src1));
+      _movp(T, Src0RM);
+      _pcmpeq(T, LEGAL_HACK(Src1RM));
     } break;
     case InstIcmp::Ne: {
-      _movp(T, Src0);
-      _pcmpeq(T, LEGAL_HACK(Src1));
+      _movp(T, Src0RM);
+      _pcmpeq(T, LEGAL_HACK(Src1RM));
       Variable *MinusOne = makeVectorOfMinusOnes(Ty);
       _pxor(T, MinusOne);
     } break;
     case InstIcmp::Ugt:
     case InstIcmp::Sgt: {
-      _movp(T, Src0);
-      _pcmpgt(T, LEGAL_HACK(Src1));
+      _movp(T, Src0RM);
+      _pcmpgt(T, LEGAL_HACK(Src1RM));
     } break;
     case InstIcmp::Uge:
     case InstIcmp::Sge: {
-      // !(Src1 > Src0)
-      _movp(T, Src1);
-      _pcmpgt(T, LEGAL_HACK(Src0));
+      // !(Src1RM > Src0RM)
+      _movp(T, Src1RM);
+      _pcmpgt(T, LEGAL_HACK(Src0RM));
       Variable *MinusOne = makeVectorOfMinusOnes(Ty);
       _pxor(T, MinusOne);
     } break;
     case InstIcmp::Ult:
     case InstIcmp::Slt: {
-      _movp(T, Src1);
-      _pcmpgt(T, LEGAL_HACK(Src0));
+      _movp(T, Src1RM);
+      _pcmpgt(T, LEGAL_HACK(Src0RM));
     } break;
     case InstIcmp::Ule:
     case InstIcmp::Sle: {
-      // !(Src0 > Src1)
-      _movp(T, Src0);
-      _pcmpgt(T, LEGAL_HACK(Src1));
+      // !(Src0RM > Src1RM)
+      _movp(T, Src0RM);
+      _pcmpgt(T, LEGAL_HACK(Src1RM));
       Variable *MinusOne = makeVectorOfMinusOnes(Ty);
       _pxor(T, MinusOne);
     } break;
     }
 #undef LEGAL_HACK
 
     _movp(Dest, T);
     eliminateNextVectorSextInstruction(Dest);
     return;
   }
@@ skipping 71 matching lines @@
   InstX8632Label *Label = InstX8632Label::create(Func, this);
   _cmp(Src0New, Src1);
   _mov(Dest, One);
   _br(getIcmp32Mapping(Inst->getCondition()), Label);
   Context.insert(InstFakeUse::create(Func, Dest));
   _mov(Dest, Zero);
   Context.insert(Label);
 }
 
 void TargetX8632::lowerInsertElement(const InstInsertElement *Inst) {
-  Operand *SourceVectOperand = Inst->getSrc(0);
-  Operand *ElementToInsert = Inst->getSrc(1);
+  Operand *SourceVectNotLegalized = Inst->getSrc(0);
+  Operand *ElementToInsertNotLegalized = Inst->getSrc(1);
   ConstantInteger *ElementIndex =
       llvm::dyn_cast<ConstantInteger>(Inst->getSrc(2));
   // Only constant indices are allowed in PNaCl IR.
   assert(ElementIndex);
   unsigned Index = ElementIndex->getValue();
-  assert(Index < typeNumElements(SourceVectOperand->getType()));
+  assert(Index < typeNumElements(SourceVectNotLegalized->getType()));
 
-  Type Ty = SourceVectOperand->getType();
+  Type Ty = SourceVectNotLegalized->getType();
   Type ElementTy = typeElementType(Ty);
   Type InVectorElementTy = getInVectorElementType(Ty);
 
   if (ElementTy == IceType_i1) {
     // Expand the element to the appropriate size for it to be inserted
     // in the vector.
     Variable *Expanded =
         Func->makeVariable(InVectorElementTy, Context.getNode());
-    InstCast *Cast =
-        InstCast::create(Func, InstCast::Zext, Expanded, ElementToInsert);
+    InstCast *Cast = InstCast::create(Func, InstCast::Zext, Expanded,
+                                      ElementToInsertNotLegalized);
     lowerCast(Cast);
-    ElementToInsert = Expanded;
+    ElementToInsertNotLegalized = Expanded;
   }
 
   if (Ty == IceType_v8i16 || Ty == IceType_v8i1 || InstructionSet >= SSE4_1) {
     // Use insertps, pinsrb, pinsrw, or pinsrd.
-    Operand *Element = legalize(ElementToInsert, Legal_Mem | Legal_Reg);
+    Operand *ElementRM =
+        legalize(ElementToInsertNotLegalized, Legal_Reg | Legal_Mem);
+    Operand *SourceVectRM =
+        legalize(SourceVectNotLegalized, Legal_Reg | Legal_Mem);
     Variable *T = makeReg(Ty);
-    _movp(T, SourceVectOperand);
+    _movp(T, SourceVectRM);
     if (Ty == IceType_v4f32)
-      _insertps(T, Element, Ctx->getConstantInt(IceType_i8, Index << 4));
+      _insertps(T, ElementRM, Ctx->getConstantInt(IceType_i8, Index << 4));
     else
-      _pinsr(T, Element, Ctx->getConstantInt(IceType_i8, Index));
+      _pinsr(T, ElementRM, Ctx->getConstantInt(IceType_i8, Index));
     _movp(Inst->getDest(), T);
   } else if (Ty == IceType_v4i32 || Ty == IceType_v4f32 || Ty == IceType_v4i1) {
     // Use shufps or movss.
-    Variable *Element = NULL;
+    Variable *ElementR = NULL;
+    Operand *SourceVectRM =
+        legalize(SourceVectNotLegalized, Legal_Reg | Legal_Mem);
+
     if (InVectorElementTy == IceType_f32) {
-      // Element will be in an XMM register since it is floating point.
-      Element = legalizeToVar(ElementToInsert);
+      // ElementR will be in an XMM register since it is floating point.
+      ElementR = legalizeToVar(ElementToInsertNotLegalized);
     } else {
       // Copy an integer to an XMM register.
-      Operand *T = legalize(ElementToInsert, Legal_Reg | Legal_Mem);
-      Element = makeReg(Ty);
-      _movd(Element, T);
+      Operand *T = legalize(ElementToInsertNotLegalized, Legal_Reg | Legal_Mem);
+      ElementR = makeReg(Ty);
+      _movd(ElementR, T);
     }
 
     if (Index == 0) {
       Variable *T = makeReg(Ty);
-      _movp(T, SourceVectOperand);
-      _movss(T, Element);
+      _movp(T, SourceVectRM);
+      _movss(T, ElementR);
       _movp(Inst->getDest(), T);
       return;
     }
 
     // shufps treats the source and desination operands as vectors of
     // four doublewords.  The destination's two high doublewords are
     // selected from the source operand and the two low doublewords are
     // selected from the (original value of) the destination operand.
     // An insertelement operation can be effected with a sequence of two
     // shufps operations with appropriate masks.  In all cases below,
     // Element[0] is being inserted into SourceVectOperand.  Indices are
     // ordered from left to right.
     //
-    // insertelement into index 1 (result is stored in Element):
-    //   Element := Element[0, 0] SourceVectOperand[0, 0]
-    //   Element := Element[3, 0] SourceVectOperand[2, 3]
+    // insertelement into index 1 (result is stored in ElementR):
+    //   ElementR := ElementR[0, 0] SourceVectRM[0, 0]
+    //   ElementR := ElementR[3, 0] SourceVectRM[2, 3]
     //
     // insertelement into index 2 (result is stored in T):
-    //   T := SourceVectOperand
-    //   Element := Element[0, 0] T[0, 3]
-    //   T := T[0, 1] Element[0, 3]
+    //   T := SourceVectRM
+    //   ElementR := ElementR[0, 0] T[0, 3]
+    //   T := T[0, 1] ElementR[0, 3]
     //
     // insertelement into index 3 (result is stored in T):
-    //   T := SourceVectOperand
-    //   Element := Element[0, 0] T[0, 2]
-    //   T := T[0, 1] Element[3, 0]
+    //   T := SourceVectRM
+    //   ElementR := ElementR[0, 0] T[0, 2]
+    //   T := T[0, 1] ElementR[3, 0]
     const unsigned char Mask1[3] = {0, 192, 128};
     const unsigned char Mask2[3] = {227, 196, 52};
 
     Constant *Mask1Constant = Ctx->getConstantInt(IceType_i8, Mask1[Index - 1]);
     Constant *Mask2Constant = Ctx->getConstantInt(IceType_i8, Mask2[Index - 1]);
 
     // ALIGNHACK: Force vector operands to registers in instructions that
     // require aligned memory operands until support for stack alignment
     // is implemented.
 #define ALIGN_HACK(Vect) legalizeToVar((Vect))
     if (Index == 1) {
-      SourceVectOperand = ALIGN_HACK(SourceVectOperand);
-      _shufps(Element, SourceVectOperand, Mask1Constant);
-      _shufps(Element, SourceVectOperand, Mask2Constant);
-      _movp(Inst->getDest(), Element);
+      SourceVectRM = ALIGN_HACK(SourceVectRM);
+      _shufps(ElementR, SourceVectRM, Mask1Constant);
+      _shufps(ElementR, SourceVectRM, Mask2Constant);
+      _movp(Inst->getDest(), ElementR);
     } else {
       Variable *T = makeReg(Ty);
-      _movp(T, SourceVectOperand);
-      _shufps(Element, T, Mask1Constant);
-      _shufps(T, Element, Mask2Constant);
+      _movp(T, SourceVectRM);
+      _shufps(ElementR, T, Mask1Constant);
+      _shufps(T, ElementR, Mask2Constant);
       _movp(Inst->getDest(), T);
     }
 #undef ALIGN_HACK
   } else {
     assert(Ty == IceType_v16i8 || Ty == IceType_v16i1);
     // Spill the value to a stack slot and perform the insertion in
     // memory.
     //
-    // TODO(wala): use legalize(SourceVectOperand, Legal_Mem) when
+    // TODO(wala): use legalize(SourceVectRM, Legal_Mem) when
     // support for legalizing to mem is implemented.
     Variable *Slot = Func->makeVariable(Ty, Context.getNode());
     Slot->setWeight(RegWeight::Zero);
-    _movp(Slot, legalizeToVar(SourceVectOperand));
+    _movp(Slot, legalizeToVar(SourceVectNotLegalized));
 
     // Compute the location of the position to insert in memory.
     unsigned Offset = Index * typeWidthInBytes(InVectorElementTy);
     OperandX8632Mem *Loc =
         getMemoryOperandForStackSlot(InVectorElementTy, Slot, Offset);
-    _store(legalizeToVar(ElementToInsert), Loc);
+    _store(legalizeToVar(ElementToInsertNotLegalized), Loc);
 
     Variable *T = makeReg(Ty);
     _movp(T, Slot);
     _movp(Inst->getDest(), T);
   }
 }
 
 void TargetX8632::lowerIntrinsicCall(const InstIntrinsicCall *Instr) {
   switch (Instr->getIntrinsicInfo().ID) {
   case Intrinsics::AtomicCmpxchg: {
@@ skipping 904 matching lines @@
 
 void TargetX8632::lowerSelect(const InstSelect *Inst) {
   Variable *Dest = Inst->getDest();
   Operand *SrcT = Inst->getTrueOperand();
   Operand *SrcF = Inst->getFalseOperand();
   Operand *Condition = Inst->getCondition();
 
   if (isVectorType(Dest->getType())) {
     Type SrcTy = SrcT->getType();
     Variable *T = makeReg(SrcTy);
+    Operand *SrcTRM = legalize(SrcT, Legal_Reg | Legal_Mem);
+    Operand *SrcFRM = legalize(SrcF, Legal_Reg | Legal_Mem);
     // ALIGNHACK: Until stack alignment support is implemented, vector
     // instructions need to have vector operands in registers.  Once
     // there is support for stack alignment, LEGAL_HACK can be removed.
 #define LEGAL_HACK(Vect) legalizeToVar((Vect))
     if (InstructionSet >= SSE4_1) {
       // TODO(wala): If the condition operand is a constant, use blendps
       // or pblendw.
       //
       // Use blendvps or pblendvb to implement select.
       if (SrcTy == IceType_v4i1 || SrcTy == IceType_v4i32 ||
           SrcTy == IceType_v4f32) {
+        Operand *ConditionRM = legalize(Condition, Legal_Reg | Legal_Mem);
         Variable *xmm0 = makeReg(IceType_v4i32, Reg_xmm0);
-        _movp(xmm0, Condition);
+        _movp(xmm0, ConditionRM);
         _psll(xmm0, Ctx->getConstantInt(IceType_i8, 31));
-        _movp(T, SrcF);
-        _blendvps(T, LEGAL_HACK(SrcT), xmm0);
+        _movp(T, SrcFRM);
+        _blendvps(T, LEGAL_HACK(SrcTRM), xmm0);
         _movp(Dest, T);
       } else {
         assert(typeNumElements(SrcTy) == 8 || typeNumElements(SrcTy) == 16);
         Type SignExtTy = Condition->getType() == IceType_v8i1 ? IceType_v8i16
-            : IceType_v16i8;
+                                                              : IceType_v16i8;
         Variable *xmm0 = makeReg(SignExtTy, Reg_xmm0);
         lowerCast(InstCast::create(Func, InstCast::Sext, xmm0, Condition));
-        _movp(T, SrcF);
-        _pblendvb(T, LEGAL_HACK(SrcT), xmm0);
+        _movp(T, SrcFRM);
+        _pblendvb(T, LEGAL_HACK(SrcTRM), xmm0);
         _movp(Dest, T);
       }
       return;
     }
     // Lower select without SSE4.1:
     // a=d?b:c ==>
     //   if elementtype(d) != i1:
     //      d=sext(d);
     //   a=(b&d)|(c&~d);
     Variable *T2 = makeReg(SrcTy);
     // Sign extend the condition operand if applicable.
     if (SrcTy == IceType_v4f32) {
       // The sext operation takes only integer arguments.
       Variable *T3 = Func->makeVariable(IceType_v4i32, Context.getNode());
       lowerCast(InstCast::create(Func, InstCast::Sext, T3, Condition));
       _movp(T, T3);
     } else if (typeElementType(SrcTy) != IceType_i1) {
       lowerCast(InstCast::create(Func, InstCast::Sext, T, Condition));
     } else {
-      _movp(T, Condition);
+      Operand *ConditionRM = legalize(Condition, Legal_Reg | Legal_Mem);
+      _movp(T, ConditionRM);
     }
     _movp(T2, T);
-    _pand(T, LEGAL_HACK(SrcT));
-    _pandn(T2, LEGAL_HACK(SrcF));
+    _pand(T, LEGAL_HACK(SrcTRM));
+    _pandn(T2, LEGAL_HACK(SrcFRM));
     _por(T, T2);
     _movp(Dest, T);
 #undef LEGAL_HACK
 
     return;
   }
 
   // a=d?b:c ==> cmp d,0; a=b; jne L1; FakeUse(a); a=c; L1:
   Operand *ConditionRMI = legalize(Condition);
   Constant *Zero = Ctx->getConstantZero(IceType_i32);
@@ skipping 305 matching lines @@
 
 OperandX8632Mem *TargetX8632::FormMemoryOperand(Operand *Operand, Type Ty) {
   OperandX8632Mem *Mem = llvm::dyn_cast<OperandX8632Mem>(Operand);
   // It may be the case that address mode optimization already creates
   // an OperandX8632Mem, so in that case it wouldn't need another level
   // of transformation.
   if (!Mem) {
     Variable *Base = llvm::dyn_cast<Variable>(Operand);
     Constant *Offset = llvm::dyn_cast<Constant>(Operand);
     assert(Base || Offset);
+    if (Offset) {
+      assert(llvm::isa<ConstantInteger>(Offset) ||
+             llvm::isa<ConstantRelocatable>(Offset));
+    }
     Mem = OperandX8632Mem::create(Func, Ty, Base, Offset);
   }
   return llvm::cast<OperandX8632Mem>(legalize(Mem));
 }
 
 Variable *TargetX8632::makeReg(Type Type, int32_t RegNum) {
   // There aren't any 64-bit integer registers for x86-32.
   assert(Type != IceType_i64);
   Variable *Reg = Func->makeVariable(Type, Context.getNode());
   if (RegNum == Variable::NoRegister)
@@ skipping 79 matching lines @@
   // It would be better to prefix with ".L$" instead of "L$", but
   // llvm-mc doesn't parse "dword ptr [.L$foo]".
   Str << "dword ptr [L$" << IceType_f32 << "$" << getPoolEntryID() << "]";
 }
 
 template <> void ConstantDouble::emit(GlobalContext *Ctx) const {
   Ostream &Str = Ctx->getStrEmit();
   Str << "qword ptr [L$" << IceType_f64 << "$" << getPoolEntryID() << "]";
 }
 
+void ConstantUndef::emit(GlobalContext *Ctx) const {
+  (void)Ctx;

[Jim Stichnoth, 2014/07/31 00:52:42]

Can you declare the function as "void ConstantUndef::emit(GlobalContext *)" and
omit this statement?

[wala, 2014/07/31 16:06:04]

Done.

+  llvm_unreachable("undef value encountered by emitter.");
+}
+
 TargetGlobalInitX8632::TargetGlobalInitX8632(GlobalContext *Ctx)
     : TargetGlobalInitLowering(Ctx) {}
 
 namespace {
 char hexdigit(unsigned X) { return X < 10 ? '0' + X : 'A' + X - 10; }
 }
 
 void TargetGlobalInitX8632::lower(const IceString &Name, SizeT Align,
                                   bool IsInternal, bool IsConst,
                                   bool IsZeroInitializer, SizeT Size,
@@ skipping 77 matching lines @@
     for (SizeT i = 0; i < Size; ++i) {
       Str << "\t.byte\t" << (((unsigned)Data[i]) & 0xff) << "\n";
     }
     Str << "\t.size\t" << MangledName << ", " << Size << "\n";
   }
   Str << "\t" << (IsInternal ? ".local" : ".global") << "\t" << MangledName
       << "\n";
 }
 
 } // end of namespace Ice