Add bfloat16

enzyme/Enzyme/CApi.cpp

@@ -99,6 +99,8 @@ ConcreteType eunwrap(CConcreteType CDT, llvm::LLVMContext &ctx) {
     return ConcreteType(llvm::Type::getDoubleTy(ctx));
   case DT_X86_FP80:
     return ConcreteType(llvm::Type::getX86_FP80Ty(ctx));
+  case DT_BFloat16:
+    return ConcreteType(llvm::Type::getBFloatTy(ctx));
   case DT_Unknown:
     return BaseType::Unknown;
   }
@@ -131,6 +133,8 @@ CConcreteType ewrap(const ConcreteType &CT) {
       return DT_Double;
     if (flt->isX86_FP80Ty())
       return DT_X86_FP80;
+    if (flt->isBFloatTy())
+      return DT_BFloat16;
   } else {
     switch (CT.SubTypeEnum) {
     case BaseType::Integer:

enzyme/Enzyme/CApi.h

@@ -60,6 +60,7 @@ typedef enum {
   DT_Double = 5,
   DT_Unknown = 6,
   DT_X86_FP80 = 7,
+  DT_BFloat16 = 8,
 } CConcreteType;
 
 struct CDataPair {

enzyme/Enzyme/InstructionDerivatives.td

@@ -251,6 +251,10 @@ def MantissaMaskOfReturnForFrexp : GlobalExpr</*primal*/0, /*shadow*/0, [{
     // x86_fp80 has only 15 exponent bits, but we must also
     // retain the most-significant bit of the mantissa as
     // there is no implicit leading
+  } else if (ty->isBFloatTy()) {
+    tsize = 16;
+    high = tsize - 1;
+    low = high - 8;
   } else if (ty->isFP128Ty()) {
     tsize = 128;
     high = tsize - 1;

enzyme/Enzyme/TypeAnalysis/ConcreteType.h

@@ -79,6 +79,8 @@ class ConcreteType {
         SubType = llvm::Type::getDoubleTy(C);
       } else if (SubName == "fp80") {
         SubType = llvm::Type::getX86_FP80Ty(C);
+      } else if (SubName == "bf16") {
+        SubType = llvm::Type::getBFloatTy(C);
       } else if (SubName == "fp128") {
         SubType = llvm::Type::getFP128Ty(C);
       } else if (SubName == "ppc128") {
@@ -104,6 +106,8 @@ class ConcreteType {
         Result += "@double";
       } else if (SubType->isX86_FP80Ty()) {
         Result += "@fp80";
+      } else if (SubType->isBFloatTy()) {
+        Result += "@bf16";
       } else if (SubType->isFP128Ty()) {
         Result += "@fp128";
       } else if (SubType->isPPC_FP128Ty()) {

enzyme/Enzyme/Utils.cpp

@@ -478,6 +478,8 @@ static inline std::string tofltstr(Type *T) {
     return "double";
   case Type::X86_FP80TyID:
     return "x87d";
+  case Type::BFloatTyID:
+    return "bf16";
   case Type::FP128TyID:
     return "quad";
   case Type::PPC_FP128TyID:

enzyme/Enzyme/Utils.h

@@ -583,6 +583,8 @@ static inline llvm::Type *FloatToIntTy(llvm::Type *T) {
   }
   if (T->isHalfTy())
     return llvm::IntegerType::get(T->getContext(), 16);
+  if (T->isBFloatTy())
+    return llvm::IntegerType::get(T->getContext(), 16);
   if (T->isFloatTy())
     return llvm::IntegerType::get(T->getContext(), 32);
   if (T->isDoubleTy())
@@ -605,6 +607,7 @@ static inline llvm::Type *IntToFloatTy(llvm::Type *T) {
     switch (ty->getBitWidth()) {
     case 16:
       return llvm::Type::getHalfTy(T->getContext());
+      // return llvm::Type::getBFloat16Ty(T->getContext());
     case 32:
       return llvm::Type::getFloatTy(T->getContext());
     case 64: