Remove a lot of the fancy scalar replacement code for dealing with llvm-gcc's

lowering of NEON code. It provides little-to-no benefit now and only introduces
additional complexity.

llvm-svn: 141646

1 files changed, 16 insertions, 205 deletions

diff --git a/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp
index b4bfdd52eb8..3a4fac07ad3 100644
--- a/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp
+++ b/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp
@@ -298,8 +298,6 @@ AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
   if (ScalarKind == Unknown)
     ScalarKind = Integer;
 
-  // FIXME: It should be possible to promote the vector type up to the alloca's
-  // size.
   if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8)
     ScalarKind = Integer;
 
@@ -334,16 +332,12 @@ AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
 /// (VectorTy) so far at the offset specified by Offset (which is specified in
 /// bytes).
 ///
-/// There are three cases we handle here:
+/// There are two cases we handle here:
 ///   1) A union of vector types of the same size and potentially its elements.
 ///      Here we turn element accesses into insert/extract element operations.
 ///      This promotes a <4 x float> with a store of float to the third element
 ///      into a <4 x float> that uses insert element.
-///   2) A union of vector types with power-of-2 size differences, e.g. a float,
-///      <2 x float> and <4 x float>.  Here we turn element accesses into insert
-///      and extract element operations, and <2 x float> accesses into a cast to
-///      <2 x double>, an extract, and a cast back to <2 x float>.
-///   3) A fully general blob of memory, which we turn into some (potentially
+///   2) A fully general blob of memory, which we turn into some (potentially
 ///      large) integer type with extract and insert operations where the loads
 ///      and stores would mutate the memory.  We mark this by setting VectorTy
 ///      to VoidTy.
@@ -374,20 +368,13 @@ void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In,
     // if the implied vector agrees with what we already have and if Offset is
     // compatible with it.
     if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
-        (!VectorTy || Offset * 8 < VectorTy->getPrimitiveSizeInBits())) {
+        (!VectorTy || EltSize == VectorTy->getElementType()
+                                         ->getPrimitiveSizeInBits()/8)) {
       if (!VectorTy) {
         ScalarKind = ImplicitVector;
         VectorTy = VectorType::get(In, AllocaSize/EltSize);
-        return;
       }
-
-      unsigned CurrentEltSize = VectorTy->getElementType()
-                                ->getPrimitiveSizeInBits()/8;
-      if (EltSize == CurrentEltSize)
-        return;
-
-      if (In->isIntegerTy() && isPowerOf2_32(AllocaSize / EltSize))
-        return;
+      return;
     }
   }
 
@@ -400,78 +387,19 @@ void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In,
 /// returning true if the type was successfully merged and false otherwise.
 bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy,
                                             uint64_t Offset) {
-  // TODO: Support nonzero offsets?
-  if (Offset != 0)
-    return false;
-
-  // Only allow vectors that are a power-of-2 away from the size of the alloca.
-  if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8)))
-    return false;
-
-  // If this the first vector we see, remember the type so that we know the
-  // element size.
-  if (!VectorTy) {
+  if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
+    // If we're storing/loading a vector of the right size, allow it as a
+    // vector.  If this the first vector we see, remember the type so that
+    // we know the element size. If this is a subsequent access, ignore it
+    // even if it is a differing type but the same size. Worst case we can
+    // bitcast the resultant vectors.
+    if (!VectorTy)
+      VectorTy = VInTy;
     ScalarKind = Vector;
-    VectorTy = VInTy;
     return true;
   }
 
-  unsigned BitWidth = VectorTy->getBitWidth();
-  unsigned InBitWidth = VInTy->getBitWidth();
-
-  // Vectors of the same size can be converted using a simple bitcast.
-  if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) {
-    ScalarKind = Vector;
-    return true;
-  }
-
-  Type *ElementTy = VectorTy->getElementType();
-  Type *InElementTy = VInTy->getElementType();
-
-  // If they're the same alloc size, we'll be attempting to convert between
-  // them with a vector shuffle, which requires the element types to match.
-  if (TD.getTypeAllocSize(VectorTy) == TD.getTypeAllocSize(VInTy) &&
-      ElementTy != InElementTy)
-    return false;
-
-  // Do not allow mixed integer and floating-point accesses from vectors of
-  // different sizes.
-  if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy())
-    return false;
-
-  if (ElementTy->isFloatingPointTy()) {
-    // Only allow floating-point vectors of different sizes if they have the
-    // same element type.
-    // TODO: This could be loosened a bit, but would anything benefit?
-    if (ElementTy != InElementTy)
-      return false;
-
-    // There are no arbitrary-precision floating-point types, which limits the
-    // number of legal vector types with larger element types that we can form
-    // to bitcast and extract a subvector.
-    // TODO: We could support some more cases with mixed fp128 and double here.
-    if (!(BitWidth == 64 || BitWidth == 128) ||
-        !(InBitWidth == 64 || InBitWidth == 128))
-      return false;
-  } else {
-    assert(ElementTy->isIntegerTy() && "Vector elements must be either integer "
-                                       "or floating-point.");
-    unsigned BitWidth = ElementTy->getPrimitiveSizeInBits();
-    unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits();
-
-    // Do not allow integer types smaller than a byte or types whose widths are
-    // not a multiple of a byte.
-    if (BitWidth < 8 || InBitWidth < 8 ||
-        BitWidth % 8 != 0 || InBitWidth % 8 != 0)
-      return false;
-  }
-
-  // Pick the largest of the two vector types.
-  ScalarKind = Vector;
-  if (InBitWidth > BitWidth)
-    VectorTy = VInTy;
-
-  return true;
+  return false;
 }
 
 /// CanConvertToScalar - V is a pointer.  If we can convert the pointee and all
@@ -735,63 +663,6 @@ void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
   }
 }
 
-/// getScaledElementType - Gets a scaled element type for a partial vector
-/// access of an alloca. The input types must be integer or floating-point
-/// scalar or vector types, and the resulting type is an integer, float or
-/// double.
-static Type *getScaledElementType(Type *Ty1, Type *Ty2,
-                                        unsigned NewBitWidth) {
-  bool IsFP1 = Ty1->isFloatingPointTy() ||
-               (Ty1->isVectorTy() &&
-                cast<VectorType>(Ty1)->getElementType()->isFloatingPointTy());
-  bool IsFP2 = Ty2->isFloatingPointTy() ||
-               (Ty2->isVectorTy() &&
-                cast<VectorType>(Ty2)->getElementType()->isFloatingPointTy());
-
-  LLVMContext &Context = Ty1->getContext();
-
-  // Prefer floating-point types over integer types, as integer types may have
-  // been created by earlier scalar replacement.
-  if (IsFP1 || IsFP2) {
-    if (NewBitWidth == 32)
-      return Type::getFloatTy(Context);
-    if (NewBitWidth == 64)
-      return Type::getDoubleTy(Context);
-  }
-
-  return Type::getIntNTy(Context, NewBitWidth);
-}
-
-/// CreateShuffleVectorCast - Creates a shuffle vector to convert one vector
-/// to another vector of the same element type which has the same allocation
-/// size but different primitive sizes (e.g. <3 x i32> and <4 x i32>).
-static Value *CreateShuffleVectorCast(Value *FromVal, Type *ToType,
-                                      IRBuilder<> &Builder) {
-  Type *FromType = FromVal->getType();
-  VectorType *FromVTy = cast<VectorType>(FromType);
-  VectorType *ToVTy = cast<VectorType>(ToType);
-  assert((ToVTy->getElementType() == FromVTy->getElementType()) &&
-         "Vectors must have the same element type");
-   Value *UnV = UndefValue::get(FromType);
-   unsigned numEltsFrom = FromVTy->getNumElements();
-   unsigned numEltsTo = ToVTy->getNumElements();
-
-   SmallVector<Constant*, 3> Args;
-   Type* Int32Ty = Builder.getInt32Ty();
-   unsigned minNumElts = std::min(numEltsFrom, numEltsTo);
-   unsigned i;
-   for (i=0; i != minNumElts; ++i)
-     Args.push_back(ConstantInt::get(Int32Ty, i));
-
-   if (i < numEltsTo) {
-     Constant* UnC = UndefValue::get(Int32Ty);
-     for (; i != numEltsTo; ++i)
-       Args.push_back(UnC);
-   }
-   Constant *Mask = ConstantVector::get(Args);
-   return Builder.CreateShuffleVector(FromVal, UnV, Mask, "tmpV");
-}
-
 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
 /// or vector value FromVal, extracting the bits from the offset specified by
 /// Offset.  This returns the value, which is of type ToType.
@@ -815,38 +686,8 @@ ConvertScalar_ExtractValue(Value *FromVal, Type *ToType,
   if (VectorType *VTy = dyn_cast<VectorType>(FromType)) {
     unsigned FromTypeSize = TD.getTypeAllocSize(FromType);
     unsigned ToTypeSize = TD.getTypeAllocSize(ToType);
-    if (FromTypeSize == ToTypeSize) {
-      // If the two types have the same primitive size, use a bit cast.
-      // Otherwise, it is two vectors with the same element type that has
-      // the same allocation size but different number of elements so use
-      // a shuffle vector.
-      if (FromType->getPrimitiveSizeInBits() ==
-          ToType->getPrimitiveSizeInBits())
+    if (FromTypeSize == ToTypeSize)
         return Builder.CreateBitCast(FromVal, ToType);
-      else
-        return CreateShuffleVectorCast(FromVal, ToType, Builder);
-    }
-
-    if (isPowerOf2_64(FromTypeSize / ToTypeSize)) {
-      assert(!(ToType->isVectorTy() && Offset != 0) && "Can't extract a value "
-             "of a smaller vector type at a nonzero offset.");
-
-      Type *CastElementTy = getScaledElementType(FromType, ToType,
-                                                       ToTypeSize * 8);
-      unsigned NumCastVectorElements = FromTypeSize / ToTypeSize;
-
-      LLVMContext &Context = FromVal->getContext();
-      Type *CastTy = VectorType::get(CastElementTy,
-                                           NumCastVectorElements);
-      Value *Cast = Builder.CreateBitCast(FromVal, CastTy);
-
-      unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy);
-      unsigned Elt = Offset/EltSize;
-      assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
-      Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get(
-                                        Type::getInt32Ty(Context), Elt));
-      return Builder.CreateBitCast(Extract, ToType);
-    }
 
     // Otherwise it must be an element access.
     unsigned Elt = 0;
@@ -961,38 +802,8 @@ ConvertScalar_InsertValue(Value *SV, Value *Old,
 
     // Changing the whole vector with memset or with an access of a different
     // vector type?
-    if (ValSize == VecSize) {
-      // If the two types have the same primitive size, use a bit cast.
-      // Otherwise, it is two vectors with the same element type that has
-      // the same allocation size but different number of elements so use
-      // a shuffle vector.
-      if (VTy->getPrimitiveSizeInBits() ==
-          SV->getType()->getPrimitiveSizeInBits())
+    if (ValSize == VecSize)
         return Builder.CreateBitCast(SV, AllocaType);
-      else
-        return CreateShuffleVectorCast(SV, VTy, Builder);
-    }
-
-    if (isPowerOf2_64(VecSize / ValSize)) {
-      assert(!(SV->getType()->isVectorTy() && Offset != 0) && "Can't insert a "
-             "value of a smaller vector type at a nonzero offset.");
-
-      Type *CastElementTy = getScaledElementType(VTy, SV->getType(),
-                                                       ValSize);
-      unsigned NumCastVectorElements = VecSize / ValSize;
-
-      Type *OldCastTy = VectorType::get(CastElementTy, NumCastVectorElements);
-      Value *OldCast = Builder.CreateBitCast(Old, OldCastTy);
-
-      Value *SVCast = Builder.CreateBitCast(SV, CastElementTy);
-
-      unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy);
-      unsigned Elt = Offset/EltSize;
-      assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
-      Value *Insert =
-        Builder.CreateInsertElement(OldCast, SVCast, Builder.getInt32(Elt));
-      return Builder.CreateBitCast(Insert, AllocaType);
-    }
 
     // Must be an element insertion.
     assert(SV->getType() == VTy->getElementType());