IR: Allow vectors of halfs to be ConstantDataVectors

Currently, vectors of halfs end up as ConstantVectors, but there isn't
a good reason they can't be ConstantDataVectors. This should save some
memory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@254991 91177308-0d34-0410-b5e6-96231b3b80d8

lib/IR/Constants.cpp

@@ -899,7 +899,9 @@ static Constant *getSequenceIfElementsMatch(Constant *C,
     else if (CI->getType()->isIntegerTy(64))
       return getIntSequenceIfElementsMatch<SequenceTy, uint64_t>(V);
   } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
-    if (CFP->getType()->isFloatTy())
+    if (CFP->getType()->isHalfTy())
+      return getFPSequenceIfElementsMatch<SequenceTy, uint16_t>(V);
+    else if (CFP->getType()->isFloatTy())
       return getFPSequenceIfElementsMatch<SequenceTy, uint32_t>(V);
     else if (CFP->getType()->isDoubleTy())
       return getFPSequenceIfElementsMatch<SequenceTy, uint64_t>(V);
@@ -2365,7 +2367,7 @@ StringRef ConstantDataSequential::getRawDataValues() const {
 /// ConstantDataArray only works with normal float and int types that are
 /// stored densely in memory, not with things like i42 or x86_f80.
 bool ConstantDataSequential::isElementTypeCompatible(Type *Ty) {
-  if (Ty->isFloatTy() || Ty->isDoubleTy()) return true;
+  if (Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy()) return true;
   if (auto *IT = dyn_cast<IntegerType>(Ty)) {
     switch (IT->getBitWidth()) {
     case 8:
@@ -2637,6 +2639,11 @@ Constant *ConstantDataVector::getSplat(unsigned NumElts, Constant *V) {
   }
 
   if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
+    if (CFP->getType()->isHalfTy()) {
+      SmallVector<uint16_t, 16> Elts(
+          NumElts, CFP->getValueAPF().bitcastToAPInt().getLimitedValue());
+      return getFP(V->getContext(), Elts);
+    }
     if (CFP->getType()->isFloatTy()) {
       SmallVector<uint32_t, 16> Elts(
           NumElts, CFP->getValueAPF().bitcastToAPInt().getLimitedValue());
@@ -2682,6 +2689,10 @@ APFloat ConstantDataSequential::getElementAsAPFloat(unsigned Elt) const {
   switch (getElementType()->getTypeID()) {
   default:
     llvm_unreachable("Accessor can only be used when element is float/double!");
+  case Type::HalfTyID: {
+    auto EltVal = *reinterpret_cast<const uint16_t *>(EltPtr);
+    return APFloat(APFloat::IEEEhalf, APInt(16, EltVal));
+  }
   case Type::FloatTyID: {
     auto EltVal = *reinterpret_cast<const uint32_t *>(EltPtr);
     return APFloat(APFloat::IEEEsingle, APInt(32, EltVal));
@@ -2716,7 +2727,8 @@ double ConstantDataSequential::getElementAsDouble(unsigned Elt) const {
 /// Note that this has to compute a new constant to return, so it isn't as
 /// efficient as getElementAsInteger/Float/Double.
 Constant *ConstantDataSequential::getElementAsConstant(unsigned Elt) const {
-  if (getElementType()->isFloatTy() || getElementType()->isDoubleTy())
+  if (getElementType()->isHalfTy() || getElementType()->isFloatTy() ||
+      getElementType()->isDoubleTy())
     return ConstantFP::get(getContext(), getElementAsAPFloat(Elt));
 
   return ConstantInt::get(getElementType(), getElementAsInteger(Elt));

test/CodeGen/X86/float-asmprint.ll

@@ -9,6 +9,8 @@
 @var64 = global double -0.0, align 8
 @var32 = global float -0.0, align 4
 @var16 = global half -0.0, align 2
+@var4f32 = global <4 x float> <float -0.0, float 0.0, float 1.0, float 2.0>
+@var4f16 = global <4 x half> <half -0.0, half 0.0, half 1.0, half 2.0>
 
 ; CHECK: var128:
 ; CHECK-NEXT: .quad 0                         # fp128 -0
@@ -39,3 +41,16 @@
 ; CHECK-NEXT: .short 32768                    # half -0
 ; CHECK-NEXT: .size
 
+; CHECK: var4f32:
+; CHECK-NEXT: .long 2147483648               # float -0
+; CHECK-NEXT: .long 0                        # float 0
+; CHECK-NEXT: .long 1065353216               # float 1
+; CHECK-NEXT: .long 1073741824               # float 2
+; CHECK-NEXT: .size
+
+; CHECK: var4f16:
+; CHECK-NEXT: .short 32768                    # half -0
+; CHECK-NEXT: .short 0                        # half 0
+; CHECK-NEXT: .short 15360                    # half 1
+; CHECK-NEXT: .short 16384                    # half 2
+; CHECK-NEXT: .size

unittests/IR/ConstantsTest.cpp

@@ -382,5 +382,33 @@ TEST(ConstantsTest, AliasCAPI) {
   ASSERT_EQ(unwrap<GlobalAlias>(AliasRef)->getAliasee(), Aliasee);
 }
 
+static std::string getNameOfType(Type *T) {
+  std::string S;
+  raw_string_ostream RSOS(S);
+  T->print(RSOS);
+  return S;
+}
+
+TEST(ConstantsTest, BuildConstantDataVectors) {
+  LLVMContext Context;
+  std::unique_ptr<Module> M(new Module("MyModule", Context));
+
+  for (Type *T : {Type::getInt8Ty(Context), Type::getInt16Ty(Context),
+                  Type::getInt32Ty(Context), Type::getInt64Ty(Context)}) {
+    Constant *Vals[] = {ConstantInt::get(T, 0), ConstantInt::get(T, 1)};
+    Constant *CDV = ConstantVector::get(Vals);
+    ASSERT_TRUE(dyn_cast<ConstantDataVector>(CDV) != nullptr)
+        << " T = " << getNameOfType(T);
+  }
+
+  for (Type *T : {Type::getHalfTy(Context), Type::getFloatTy(Context),
+                  Type::getDoubleTy(Context)}) {
+    Constant *Vals[] = {ConstantFP::get(T, 0), ConstantFP::get(T, 1)};
+    Constant *CDV = ConstantVector::get(Vals);
+    ASSERT_TRUE(dyn_cast<ConstantDataVector>(CDV) != nullptr)
+        << " T = " << getNameOfType(T);
+  }
+}
+
 }  // end anonymous namespace
 }  // end namespace llvm