[IR] Add llvm.sincos intrinsic (llvm#109825)

This adds the `llvm.sincos` intrinsic, legalization, and lowering.

The `llvm.sincos` intrinsic takes a floating-point value and returns
both the sine and cosine (as a struct).

```
declare { float, float }          @llvm.sincos.f32(float  %Val)
declare { double, double }        @llvm.sincos.f64(double %Val)
declare { x86_fp80, x86_fp80 }    @llvm.sincos.f80(x86_fp80  %Val)
declare { fp128, fp128 }          @llvm.sincos.f128(fp128 %Val)
declare { ppc_fp128, ppc_fp128 }  @llvm.sincos.ppcf128(ppc_fp128  %Val)
declare { <4 x float>, <4 x float> } @llvm.sincos.v4f32(<4 x float>  %Val)
```

The lowering is built on top of the existing FSINCOS ISD node, with
additional type legalization to allow for f16, f128, and vector values.

Author: MacDue

llvm/docs/GlobalISel/GenericOpcode.rst

@@ -633,8 +633,8 @@ G_FCEIL, G_FSQRT, G_FFLOOR, G_FRINT, G_FNEARBYINT
 
 These correspond to the standard C functions of the same name.
 
-G_FCOS, G_FSIN, G_FTAN, G_FACOS, G_FASIN, G_FATAN, G_FATAN2, G_FCOSH, G_FSINH, G_FTANH
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+G_FCOS, G_FSIN, G_FSINCOS, G_FTAN, G_FACOS, G_FASIN, G_FATAN, G_FATAN2, G_FCOSH, G_FSINH, G_FTANH
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 These correspond to the standard C trigonometry functions of the same name.
 

llvm/docs/LangRef.rst

@@ -15512,6 +15512,8 @@ Semantics:
 This function returns the first value raised to the second power with an
 unspecified sequence of rounding operations.
 
+.. _t_llvm_sin:
+
 '``llvm.sin.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^
 
@@ -15549,6 +15551,8 @@ trapping or setting ``errno``.
 When specified with the fast-math-flag 'afn', the result may be approximated
 using a less accurate calculation.
 
+.. _t_llvm_cos:
+
 '``llvm.cos.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^
 
@@ -15882,6 +15886,50 @@ trapping or setting ``errno``.
 When specified with the fast-math-flag 'afn', the result may be approximated
 using a less accurate calculation.
 
+
+'``llvm.sincos.*``' Intrinsic
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Syntax:
+"""""""
+
+This is an overloaded intrinsic. You can use ``llvm.sincos`` on any
+floating-point or vector of floating-point type. Not all targets support
+all types however.
+
+::
+
+      declare { float, float }          @llvm.sincos.f32(float  %Val)
+      declare { double, double }        @llvm.sincos.f64(double %Val)
+      declare { x86_fp80, x86_fp80 }    @llvm.sincos.f80(x86_fp80  %Val)
+      declare { fp128, fp128 }          @llvm.sincos.f128(fp128 %Val)
+      declare { ppc_fp128, ppc_fp128 }  @llvm.sincos.ppcf128(ppc_fp128  %Val)
+      declare { <4 x float>, <4 x float> } @llvm.sincos.v4f32(<4 x float>  %Val)
+
+Overview:
+"""""""""
+
+The '``llvm.sincos.*``' intrinsics returns the sine and cosine of the operand.
+
+Arguments:
+""""""""""
+
+The argument is a :ref:`floating-point <t_floating>` value or
+:ref:`vector <t_vector>` of floating-point values. Returns two values matching
+the argument type in a struct.
+
+Semantics:
+""""""""""
+
+This intrinsic is equivalent to a calling both :ref:`llvm.sin <t_llvm_sin>`
+and :ref:`llvm.cos <t_llvm_cos>` on the argument.
+
+The first result is the sine of the argument and the second result is the cosine
+of the argument.
+
+When specified with the fast-math-flag 'afn', the result may be approximated
+using a less accurate calculation.
+
 '``llvm.pow.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^
 

llvm/include/llvm/CodeGen/BasicTTIImpl.h

@@ -1986,6 +1986,9 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
     case Intrinsic::cos:
       ISD = ISD::FCOS;
       break;
+    case Intrinsic::sincos:
+      ISD = ISD::FSINCOS;
+      break;
     case Intrinsic::tan:
       ISD = ISD::FTAN;
       break;

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

@@ -2009,6 +2009,13 @@ class MachineIRBuilder {
     return buildInstr(TargetOpcode::G_FFREXP, {Fract, Exp}, {Src}, Flags);
   }
 
+  /// Build and insert \p Sin, \p Cos = G_FSINCOS \p Src
+  MachineInstrBuilder
+  buildFSincos(const DstOp &Sin, const DstOp &Cos, const SrcOp &Src,
+               std::optional<unsigned> Flags = std::nullopt) {
+    return buildInstr(TargetOpcode::G_FSINCOS, {Sin, Cos}, {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/IR/Intrinsics.td

@@ -1050,6 +1050,8 @@ let IntrProperties = [IntrNoMem, IntrSpeculatable, IntrWillReturn] in {
   def int_nearbyint : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>]>;
   def int_round : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>]>;
   def int_roundeven    : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>]>;
+  def int_sincos : DefaultAttrsIntrinsic<[LLVMMatchType<0>, LLVMMatchType<0>],
+                             [llvm_anyfloat_ty]>;
 
   // Truncate a floating point number with a specific rounding mode
   def int_fptrunc_round : DefaultAttrsIntrinsic<[ llvm_anyfloat_ty ],

llvm/include/llvm/Support/TargetOpcodes.def

@@ -809,6 +809,9 @@ HANDLE_TARGET_OPCODE(G_FCOS)
 /// Floating point sine.
 HANDLE_TARGET_OPCODE(G_FSIN)
 
+/// Floating point combined sine and cosine.
+HANDLE_TARGET_OPCODE(G_FSINCOS)
+
 /// Floating point tangent.
 HANDLE_TARGET_OPCODE(G_FTAN)
 

llvm/include/llvm/Target/GenericOpcodes.td

@@ -1020,6 +1020,13 @@ def G_FSIN : GenericInstruction {
   let hasSideEffects = false;
 }
 
+// Floating point combined sine and cosine.
+def G_FSINCOS : GenericInstruction {
+  let OutOperandList = (outs type0:$dst1, type0:$dst2);
+  let InOperandList = (ins type0:$src1);
+  let hasSideEffects = false;
+}
+
 // Floating point tangent of a value.
 def G_FTAN : GenericInstruction {
   let OutOperandList = (outs type0:$dst);

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

@@ -2343,6 +2343,13 @@ bool IRTranslator::translateKnownIntrinsic(const CallInst &CI, Intrinsic::ID ID,
                            MachineInstr::copyFlagsFromInstruction(CI));
     return true;
   }
+  case Intrinsic::sincos: {
+    ArrayRef<Register> VRegs = getOrCreateVRegs(CI);
+    MIRBuilder.buildFSincos(VRegs[0], VRegs[1],
+                            getOrCreateVReg(*CI.getArgOperand(0)),
+                            MachineInstr::copyFlagsFromInstruction(CI));
+    return true;
+  }
   case Intrinsic::fptosi_sat:
     MIRBuilder.buildFPTOSI_SAT(getOrCreateVReg(CI),
                                getOrCreateVReg(*CI.getArgOperand(0)));

llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

@@ -3714,6 +3714,17 @@ bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
     }
     break;
   }
+  case ISD::FSINCOS: {
+    if (isSinCosLibcallAvailable(Node, TLI))
+      break;
+    EVT VT = Node->getValueType(0);
+    SDValue Op = Node->getOperand(0);
+    SDNodeFlags Flags = Node->getFlags();
+    Tmp1 = DAG.getNode(ISD::FSIN, dl, VT, Op, Flags);
+    Tmp2 = DAG.getNode(ISD::FCOS, dl, VT, Op, Flags);
+    Results.append({Tmp1, Tmp2});
+    break;
+  }
   case ISD::FMAD:
     llvm_unreachable("Illegal fmad should never be formed");
 
@@ -5586,6 +5597,16 @@ void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
     Results.push_back(Tmp2.getValue(1));
     break;
   }
+  case ISD::FSINCOS: {
+    Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
+    Tmp2 = DAG.getNode(ISD::FSINCOS, dl, DAG.getVTList(NVT, NVT), Tmp1,
+                       Node->getFlags());
+    Tmp3 = DAG.getIntPtrConstant(0, dl, /*isTarget=*/true);
+    for (unsigned ResNum = 0; ResNum < Node->getNumValues(); ResNum++)
+      Results.push_back(
+          DAG.getNode(ISD::FP_ROUND, dl, OVT, Tmp2.getValue(ResNum), Tmp3));
+    break;
+  }
   case ISD::FFLOOR:
   case ISD::FCEIL:
   case ISD::FRINT:

llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp

@@ -129,6 +129,7 @@ void DAGTypeLegalizer::SoftenFloatResult(SDNode *N, unsigned ResNo) {
     case ISD::FLDEXP:
     case ISD::STRICT_FLDEXP: R = SoftenFloatRes_ExpOp(N); break;
     case ISD::FFREXP:        R = SoftenFloatRes_FFREXP(N); break;
+    case ISD::FSINCOS:       R = SoftenFloatRes_FSINCOS(N); break;
     case ISD::STRICT_FREM:
     case ISD::FREM:        R = SoftenFloatRes_FREM(N); break;
     case ISD::STRICT_FRINT:
@@ -774,6 +775,45 @@ SDValue DAGTypeLegalizer::SoftenFloatRes_FFREXP(SDNode *N) {
   return ReturnVal;
 }
 
+SDValue DAGTypeLegalizer::SoftenFloatRes_FSINCOS(SDNode *N) {
+  assert(!N->isStrictFPOpcode() && "strictfp not implemented for fsincos");
+  EVT VT = N->getValueType(0);
+  RTLIB::Libcall LC = RTLIB::getFSINCOS(VT);
+
+  if (!TLI.getLibcallName(LC))
+    return SDValue();
+
+  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
+  SDValue StackSlotSin = DAG.CreateStackTemporary(NVT);
+  SDValue StackSlotCos = DAG.CreateStackTemporary(NVT);
+
+  SDLoc DL(N);
+
+  TargetLowering::MakeLibCallOptions CallOptions;
+  std::array Ops{GetSoftenedFloat(N->getOperand(0)), StackSlotSin,
+                 StackSlotCos};
+  std::array OpsVT{VT, StackSlotSin.getValueType(),
+                   StackSlotCos.getValueType()};
+
+  // TODO: setTypeListBeforeSoften can't properly express multiple return types,
+  // but since both returns have the same type for sincos it should be okay.
+  CallOptions.setTypeListBeforeSoften({OpsVT}, VT, true);
+
+  auto [ReturnVal, Chain] = TLI.makeLibCall(DAG, LC, NVT, Ops, CallOptions, DL,
+                                            /*Chain=*/SDValue());
+
+  auto CreateStackLoad = [&, Chain = Chain](SDValue StackSlot) {
+    int FrameIdx = cast<FrameIndexSDNode>(StackSlot)->getIndex();
+    auto PtrInfo =
+        MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx);
+    return DAG.getLoad(NVT, DL, Chain, StackSlot, PtrInfo);
+  };
+  SetSoftenedFloat(SDValue(N, 0), CreateStackLoad(StackSlotSin));
+  SetSoftenedFloat(SDValue(N, 1), CreateStackLoad(StackSlotCos));
+
+  return SDValue();
+}
+
 SDValue DAGTypeLegalizer::SoftenFloatRes_FREM(SDNode *N) {
   return SoftenFloatRes_Binary(N, GetFPLibCall(N->getValueType(0),
                                                RTLIB::REM_F32,
@@ -2704,6 +2744,10 @@ void DAGTypeLegalizer::PromoteFloatResult(SDNode *N, unsigned ResNo) {
     case ISD::FLDEXP:     R = PromoteFloatRes_ExpOp(N); break;
     case ISD::FFREXP:     R = PromoteFloatRes_FFREXP(N); break;
 
+    case ISD::FSINCOS:
+      R = PromoteFloatRes_FSINCOS(N);
+      break;
+
     case ISD::FP_ROUND:   R = PromoteFloatRes_FP_ROUND(N); break;
     case ISD::STRICT_FP_ROUND:
       R = PromoteFloatRes_STRICT_FP_ROUND(N);
@@ -2899,6 +2943,20 @@ SDValue DAGTypeLegalizer::PromoteFloatRes_FFREXP(SDNode *N) {
   return Res;
 }
 
+SDValue DAGTypeLegalizer::PromoteFloatRes_FSINCOS(SDNode *N) {
+  EVT VT = N->getValueType(0);
+  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
+  SDValue Op = GetPromotedFloat(N->getOperand(0));
+  SDValue Res = DAG.getNode(N->getOpcode(), SDLoc(N), {NVT, NVT}, Op);
+
+  for (unsigned ResNum = 0, NumValues = N->getNumValues(); ResNum < NumValues;
+       ++ResNum) {
+    SetPromotedFloat(SDValue(N, ResNum), Res.getValue(ResNum));
+  }
+
+  return SDValue();
+}
+
 // Explicit operation to reduce precision.  Reduce the value to half precision
 // and promote it back to the legal type.
 SDValue DAGTypeLegalizer::PromoteFloatRes_FP_ROUND(SDNode *N) {
@@ -3148,6 +3206,10 @@ void DAGTypeLegalizer::SoftPromoteHalfResult(SDNode *N, unsigned ResNo) {
 
   case ISD::FFREXP:      R = SoftPromoteHalfRes_FFREXP(N); break;
 
+  case ISD::FSINCOS:
+    R = SoftPromoteHalfRes_FSINCOS(N);
+    break;
+
   case ISD::LOAD:        R = SoftPromoteHalfRes_LOAD(N); break;
   case ISD::ATOMIC_LOAD:
     R = SoftPromoteHalfRes_ATOMIC_LOAD(N);
@@ -3304,6 +3366,27 @@ SDValue DAGTypeLegalizer::SoftPromoteHalfRes_FFREXP(SDNode *N) {
   return DAG.getNode(GetPromotionOpcode(NVT, OVT), dl, MVT::i16, Res);
 }
 
+SDValue DAGTypeLegalizer::SoftPromoteHalfRes_FSINCOS(SDNode *N) {
+  EVT OVT = N->getValueType(0);
+  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), OVT);
+  SDValue Op = GetSoftPromotedHalf(N->getOperand(0));
+  SDLoc dl(N);
+
+  // Promote to the larger FP type.
+  Op = DAG.getNode(GetPromotionOpcode(OVT, NVT), dl, NVT, Op);
+  SDValue Res = DAG.getNode(N->getOpcode(), dl, DAG.getVTList(NVT, NVT), Op);
+
+  // Convert back to FP16 as an integer.
+  ISD::NodeType Truncate = GetPromotionOpcode(NVT, OVT);
+  for (unsigned ResNum = 0, NumValues = N->getNumValues(); ResNum < NumValues;
+       ++ResNum) {
+    SDValue Trunc = DAG.getNode(Truncate, dl, MVT::i16, Res.getValue(ResNum));
+    SetSoftPromotedHalf(SDValue(N, ResNum), Trunc);
+  }
+
+  return SDValue();
+}
+
 SDValue DAGTypeLegalizer::SoftPromoteHalfRes_FP_ROUND(SDNode *N) {
   EVT RVT = N->getValueType(0);
   bool IsStrict = N->isStrictFPOpcode();