re-enable bfloat

Author: tgymnich

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

@@ -295,21 +295,8 @@ void IRTranslator::addMachineCFGPred(CFGEdge Edge, MachineBasicBlock *NewPred) {
   MachinePreds[Edge].push_back(NewPred);
 }
 
-static bool containsBF16Type(const User &U) {
-  // BF16 cannot currently be represented by LLT, to avoid miscompiles we
-  // prevent any instructions using them. FIXME: This can be removed once LLT
-  // supports bfloat.
-  return U.getType()->getScalarType()->isBFloatTy() ||
-         any_of(U.operands(), [](Value *V) {
-           return V->getType()->getScalarType()->isBFloatTy();
-         });
-}
-
 bool IRTranslator::translateBinaryOp(unsigned Opcode, const User &U,
                                      MachineIRBuilder &MIRBuilder) {
-  if (containsBF16Type(U))
-    return false;
-
   // Get or create a virtual register for each value.
   // Unless the value is a Constant => loadimm cst?
   // or inline constant each time?
@@ -329,9 +316,6 @@ bool IRTranslator::translateBinaryOp(unsigned Opcode, const User &U,
 
 bool IRTranslator::translateUnaryOp(unsigned Opcode, const User &U,
                                     MachineIRBuilder &MIRBuilder) {
-  if (containsBF16Type(U))
-    return false;
-
   Register Op0 = getOrCreateVReg(*U.getOperand(0));
   Register Res = getOrCreateVReg(U);
   uint32_t Flags = 0;
@@ -349,9 +333,6 @@ bool IRTranslator::translateFNeg(const User &U, MachineIRBuilder &MIRBuilder) {
 
 bool IRTranslator::translateCompare(const User &U,
                                     MachineIRBuilder &MIRBuilder) {
-  if (containsBF16Type(U))
-    return false;
-
   auto *CI = cast<CmpInst>(&U);
   Register Op0 = getOrCreateVReg(*U.getOperand(0));
   Register Op1 = getOrCreateVReg(*U.getOperand(1));
@@ -1571,9 +1552,6 @@ bool IRTranslator::translateBitCast(const User &U,
 
 bool IRTranslator::translateCast(unsigned Opcode, const User &U,
                                  MachineIRBuilder &MIRBuilder) {
-  if (containsBF16Type(U))
-    return false;
-
   uint32_t Flags = 0;
   if (const Instruction *I = dyn_cast<Instruction>(&U))
     Flags = MachineInstr::copyFlagsFromInstruction(*I);
@@ -2662,9 +2640,6 @@ bool IRTranslator::translateKnownIntrinsic(const CallInst &CI, Intrinsic::ID ID,
 
 bool IRTranslator::translateInlineAsm(const CallBase &CB,
                                       MachineIRBuilder &MIRBuilder) {
-  if (containsBF16Type(CB))
-    return false;
-
   const InlineAsmLowering *ALI = MF->getSubtarget().getInlineAsmLowering();
 
   if (!ALI) {
@@ -2753,9 +2728,6 @@ bool IRTranslator::translateCallBase(const CallBase &CB,
 }
 
 bool IRTranslator::translateCall(const User &U, MachineIRBuilder &MIRBuilder) {
-  if (containsBF16Type(U))
-    return false;
-
   const CallInst &CI = cast<CallInst>(U);
   const Function *F = CI.getCalledFunction();
 
@@ -3387,9 +3359,6 @@ bool IRTranslator::translateAtomicCmpXchg(const User &U,
 
 bool IRTranslator::translateAtomicRMW(const User &U,
                                       MachineIRBuilder &MIRBuilder) {
-  if (containsBF16Type(U))
-    return false;
-
   const AtomicRMWInst &I = cast<AtomicRMWInst>(U);
   auto Flags = TLI->getAtomicMemOperandFlags(I, *DL);