[AArch64][GlobalISel] Add G_FMODF instruction (llvm#160061)

This commit adds the intrinsic `G_FMODF` to GMIR & enables its
translation, legalization and instruction selection in AArch64.

Author: HolyMolyCowMan

llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h

@@ -300,6 +300,10 @@ class LegalizerHelper {
                                    Type *OpType,
                                    LostDebugLocObserver &LocObserver);
 
+  LegalizeResult emitModfLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder,
+                                 unsigned Size, Type *OpType,
+                                 LostDebugLocObserver &LocObserver);
+
 public:
   /// Return the alignment to use for a stack temporary object with the given
   /// type.

llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h

@@ -2184,6 +2184,13 @@ class LLVM_ABI MachineIRBuilder {
     return buildInstr(TargetOpcode::G_FSINCOS, {Sin, Cos}, {Src}, Flags);
   }
 
+  /// Build and insert \p Fract, \p Int = G_FMODF \p Src
+  MachineInstrBuilder buildModf(const DstOp &Fract, const DstOp &Int,
+                                const SrcOp &Src,
+                                std::optional<unsigned> Flags = std::nullopt) {
+    return buildInstr(TargetOpcode::G_FMODF, {Fract, Int}, {Src}, Flags);
+  }
+
   /// Build and insert \p Res = G_FCOPYSIGN \p Op0, \p Op1
   MachineInstrBuilder buildFCopysign(const DstOp &Dst, const SrcOp &Src0,
                                      const SrcOp &Src1) {

llvm/include/llvm/Support/TargetOpcodes.def

@@ -650,6 +650,9 @@ HANDLE_TARGET_OPCODE(G_FDIV)
 /// Generic FP remainder.
 HANDLE_TARGET_OPCODE(G_FREM)
 
+/// Generic FP modf
+HANDLE_TARGET_OPCODE(G_FMODF)
+
 /// Generic FP exponentiation.
 HANDLE_TARGET_OPCODE(G_FPOW)
 

llvm/include/llvm/Target/GenericOpcodes.td

@@ -981,6 +981,13 @@ def G_FREM : GenericInstruction {
   let hasSideEffects = false;
 }
 
+/// Generic FP modf
+def G_FMODF : GenericInstruction {
+  let OutOperandList = (outs type0:$dst1, type0:$dst2);
+  let InOperandList = (ins type0:$src1);
+  let hasSideEffects = false;
+}
+
 // Floating point exponentiation.
 def G_FPOW : GenericInstruction {
   let OutOperandList = (outs type0:$dst);

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

@@ -2362,6 +2362,13 @@ bool IRTranslator::translateKnownIntrinsic(const CallInst &CI, Intrinsic::ID ID,
                            MachineInstr::copyFlagsFromInstruction(CI));
     return true;
   }
+  case Intrinsic::modf: {
+    ArrayRef<Register> VRegs = getOrCreateVRegs(CI);
+    MIRBuilder.buildModf(VRegs[0], VRegs[1],
+                         getOrCreateVReg(*CI.getArgOperand(0)),
+                         MachineInstr::copyFlagsFromInstruction(CI));
+    return true;
+  }
   case Intrinsic::sincos: {
     ArrayRef<Register> VRegs = getOrCreateVRegs(CI);
     MIRBuilder.buildFSincos(VRegs[0], VRegs[1],

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

@@ -471,6 +471,8 @@ static RTLIB::Libcall getRTLibDesc(unsigned Opcode, unsigned Size) {
     RTLIBCASE(TANH_F);
   case TargetOpcode::G_FSINCOS:
     RTLIBCASE(SINCOS_F);
+  case TargetOpcode::G_FMODF:
+    RTLIBCASE(MODF_F);
   case TargetOpcode::G_FLOG10:
     RTLIBCASE(LOG10_F);
   case TargetOpcode::G_FLOG:
@@ -702,6 +704,46 @@ LegalizerHelper::LegalizeResult LegalizerHelper::emitSincosLibcall(
   return LegalizerHelper::Legalized;
 }
 
+LegalizerHelper::LegalizeResult
+LegalizerHelper::emitModfLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder,
+                                 unsigned Size, Type *OpType,
+                                 LostDebugLocObserver &LocObserver) {
+  MachineFunction &MF = MIRBuilder.getMF();
+  MachineRegisterInfo &MRI = MF.getRegInfo();
+
+  Register DstFrac = MI.getOperand(0).getReg();
+  Register DstInt = MI.getOperand(1).getReg();
+  Register Src = MI.getOperand(2).getReg();
+  LLT DstTy = MRI.getType(DstFrac);
+
+  int MemSize = DstTy.getSizeInBytes();
+  Align Alignment = getStackTemporaryAlignment(DstTy);
+  const DataLayout &DL = MIRBuilder.getDataLayout();
+  unsigned AddrSpace = DL.getAllocaAddrSpace();
+  MachinePointerInfo PtrInfo;
+
+  Register StackPtrInt =
+      createStackTemporary(TypeSize::getFixed(MemSize), Alignment, PtrInfo)
+          .getReg(0);
+
+  auto &Ctx = MF.getFunction().getContext();
+  auto LibcallResult = createLibcall(
+      MIRBuilder, getRTLibDesc(MI.getOpcode(), Size), {DstFrac, OpType, 0},
+      {{Src, OpType, 0}, {StackPtrInt, PointerType::get(Ctx, AddrSpace), 1}},
+      LocObserver, &MI);
+
+  if (LibcallResult != LegalizeResult::Legalized)
+    return LegalizerHelper::UnableToLegalize;
+
+  MachineMemOperand *LoadMMOInt = MF.getMachineMemOperand(
+      PtrInfo, MachineMemOperand::MOLoad, MemSize, Alignment);
+
+  MIRBuilder.buildLoad(DstInt, StackPtrInt, *LoadMMOInt);
+  MI.eraseFromParent();
+
+  return LegalizerHelper::Legalized;
+}
+
 LegalizerHelper::LegalizeResult
 llvm::createMemLibcall(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI,
                        MachineInstr &MI, LostDebugLocObserver &LocObserver) {
@@ -1341,6 +1383,16 @@ LegalizerHelper::libcall(MachineInstr &MI, LostDebugLocObserver &LocObserver) {
     }
     return emitSincosLibcall(MI, MIRBuilder, Size, HLTy, LocObserver);
   }
+  case TargetOpcode::G_FMODF: {
+    LLT LLTy = MRI.getType(MI.getOperand(0).getReg());
+    unsigned Size = LLTy.getSizeInBits();
+    Type *HLTy = getFloatTypeForLLT(Ctx, LLTy);
+    if (!HLTy || (Size != 32 && Size != 64 && Size != 80 && Size != 128)) {
+      LLVM_DEBUG(dbgs() << "No libcall available for type " << LLTy << ".\n");
+      return UnableToLegalize;
+    }
+    return emitModfLibcall(MI, MIRBuilder, Size, HLTy, LocObserver);
+  }
   case TargetOpcode::G_LROUND:
   case TargetOpcode::G_LLROUND:
   case TargetOpcode::G_INTRINSIC_LRINT:
@@ -3333,6 +3385,16 @@ LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
     widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
     Observer.changedInstr(MI);
     return Legalized;
+  case TargetOpcode::G_FMODF: {
+    Observer.changingInstr(MI);
+    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_FPEXT);
+
+    widenScalarDst(MI, WideTy, 1, TargetOpcode::G_FPTRUNC);
+    MIRBuilder.setInsertPt(MIRBuilder.getMBB(), --MIRBuilder.getInsertPt());
+    widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
+    Observer.changedInstr(MI);
+    return Legalized;
+  }
   case TargetOpcode::G_FPOWI:
   case TargetOpcode::G_FLDEXP:
   case TargetOpcode::G_STRICT_FLDEXP: {
@@ -5472,6 +5534,7 @@ LegalizerHelper::fewerElementsVector(MachineInstr &MI, unsigned TypeIdx,
   case G_LROUND:
   case G_LLROUND:
   case G_INTRINSIC_TRUNC:
+  case G_FMODF:
   case G_FCOS:
   case G_FSIN:
   case G_FTAN:

llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp

@@ -438,7 +438,7 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
   getActionDefinitionsBuilder({G_FCOS, G_FSIN, G_FPOW, G_FLOG, G_FLOG2,
                                G_FLOG10, G_FTAN, G_FEXP, G_FEXP2, G_FEXP10,
                                G_FACOS, G_FASIN, G_FATAN, G_FATAN2, G_FCOSH,
-                               G_FSINH, G_FTANH})
+                               G_FSINH, G_FTANH, G_FMODF})
       // We need a call for these, so we always need to scalarize.
       .scalarize(0)
       // Regardless of FP16 support, widen 16-bit elements to 32-bits.

llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp

@@ -752,6 +752,8 @@ bool SPIRVInstructionSelector::spvSelect(Register ResVReg,
     return selectExtInst(ResVReg, ResType, I, CL::exp, GL::Exp);
   case TargetOpcode::G_FEXP2:
     return selectExtInst(ResVReg, ResType, I, CL::exp2, GL::Exp2);
+  case TargetOpcode::G_FMODF:
+    return selectModf(ResVReg, ResType, I);
 
   case TargetOpcode::G_FLOG:
     return selectExtInst(ResVReg, ResType, I, CL::log, GL::Log);
@@ -3453,9 +3455,6 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
   case Intrinsic::spv_discard: {
     return selectDiscard(ResVReg, ResType, I);
   }
-  case Intrinsic::modf: {
-    return selectModf(ResVReg, ResType, I);
-  }
   default: {
     std::string DiagMsg;
     raw_string_ostream OS(DiagMsg);
@@ -4268,6 +4267,7 @@ bool SPIRVInstructionSelector::selectModf(Register ResVReg,
         PtrTyReg,
         LLT::pointer(storageClassToAddressSpace(SPIRV::StorageClass::Function),
                      GR.getPointerSize()));
+
     // Assign SPIR-V type of the pointer type of the alloca variable to the
     // new register.
     GR.assignSPIRVTypeToVReg(PtrType, PtrTyReg, MIRBuilder.getMF());
@@ -4280,10 +4280,7 @@ bool SPIRVInstructionSelector::selectModf(Register ResVReg,
             .addUse(GR.getSPIRVTypeID(PtrType))
             .addImm(static_cast<uint32_t>(SPIRV::StorageClass::Function));
     Register Variable = AllocaMIB->getOperand(0).getReg();
-    // Modf must have 4 operands, the first two are the 2 parts of the result,
-    // the third is the operand, and the last one is the floating point value.
-    assert(I.getNumOperands() == 4 &&
-           "Expected 4 operands for modf instruction");
+
     MachineBasicBlock &BB = *I.getParent();
     // Create the OpenCLLIB::modf instruction.
     auto MIB =
@@ -4293,8 +4290,8 @@ bool SPIRVInstructionSelector::selectModf(Register ResVReg,
             .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::OpenCL_std))
             .addImm(CL::modf)
             .setMIFlags(I.getFlags())
-            .add(I.getOperand(3)) // Floating point value.
-            .addUse(Variable);    // Pointer to integral part.
+            .add(I.getOperand(I.getNumExplicitDefs())) // Floating point value.
+            .addUse(Variable); // Pointer to integral part.
     // Assign the integral part stored in the ptr to the second element of the
     // result.
     Register IntegralPartReg = I.getOperand(1).getReg();

llvm/lib/Target/SPIRV/SPIRVLegalizerInfo.cpp

@@ -300,6 +300,7 @@ SPIRVLegalizerInfo::SPIRVLegalizerInfo(const SPIRVSubtarget &ST) {
   getActionDefinitionsBuilder({G_STRICT_FSQRT,
                                G_FPOW,
                                G_FEXP,
+                               G_FMODF,
                                G_FEXP2,
                                G_FLOG,
                                G_FLOG2,