(release 20.x non-upstream patch) [LV] Teach the vectorizer to cost and vectorize llvm.sincos intrinsics

llvm/include/llvm/Analysis/TargetTransformInfo.h

@@ -126,12 +126,17 @@ class IntrinsicCostAttributes {
   // If ScalarizationCost is UINT_MAX, the cost of scalarizing the
   // arguments and the return value will be computed based on types.
   InstructionCost ScalarizationCost = InstructionCost::getInvalid();
+  /* Downstream change: #87 (sincos vectorization)*/
+  TargetLibraryInfo const *LibInfo = nullptr;
+  /* End downstream change: #87 */
 
 public:
+  /* Downstream change: #87 (sincos vectorization)*/
   IntrinsicCostAttributes(
       Intrinsic::ID Id, const CallBase &CI,
       InstructionCost ScalarCost = InstructionCost::getInvalid(),
-      bool TypeBasedOnly = false);
+      bool TypeBasedOnly = false, TargetLibraryInfo const *LibInfo = nullptr);
+  /* End downstream change: #87 */
 
   IntrinsicCostAttributes(
       Intrinsic::ID Id, Type *RTy, ArrayRef<Type *> Tys,
@@ -141,11 +146,14 @@ class IntrinsicCostAttributes {
   IntrinsicCostAttributes(Intrinsic::ID Id, Type *RTy,
                           ArrayRef<const Value *> Args);
 
+  /* Downstream change: #87 (sincos vectorization)*/
   IntrinsicCostAttributes(
       Intrinsic::ID Id, Type *RTy, ArrayRef<const Value *> Args,
       ArrayRef<Type *> Tys, FastMathFlags Flags = FastMathFlags(),
       const IntrinsicInst *I = nullptr,
-      InstructionCost ScalarCost = InstructionCost::getInvalid());
+      InstructionCost ScalarCost = InstructionCost::getInvalid(),
+      TargetLibraryInfo const *LibInfo = nullptr);
+  /* End downstream change: #87 */
 
   Intrinsic::ID getID() const { return IID; }
   const IntrinsicInst *getInst() const { return II; }
@@ -155,6 +163,10 @@ class IntrinsicCostAttributes {
   const SmallVectorImpl<const Value *> &getArgs() const { return Arguments; }
   const SmallVectorImpl<Type *> &getArgTypes() const { return ParamTys; }
 
+  /* Downstream change: #87 (sincos vectorization)*/
+  const TargetLibraryInfo *getLibInfo() const { return LibInfo; }
+  /* End downstream change: #87 */
+
   bool isTypeBasedOnly() const {
     return Arguments.empty();
   }

llvm/include/llvm/CodeGen/BasicTTIImpl.h

@@ -22,6 +22,7 @@
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/Analysis/LoopInfo.h"
 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
+#include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/Analysis/TargetTransformInfoImpl.h"
 #include "llvm/Analysis/ValueTracking.h"
@@ -285,6 +286,66 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
     return false;
   }
 
+  /* Downstream change: #87 (sincos vectorization)*/
+  /// Several intrinsics that return structs (including llvm.sincos[pi] and
+  /// llvm.modf) can be lowered to a vector library call (for certain VFs). The
+  /// vector library functions correspond to the scalar calls (e.g. sincos or
+  /// modf), which unlike the intrinsic return values via output pointers. This
+  /// helper checks if a vector call exists for the given intrinsic, and returns
+  /// the cost, which includes the cost of the mask (if required), and the loads
+  /// for values returned via output pointers. \p LC is the scalar libcall and
+  /// \p CallRetElementIndex (optional) is the struct element which is mapped to
+  /// the call return value. If std::nullopt is returned, then no vector library
+  /// call is available, so the intrinsic should be assigned the default cost
+  /// (e.g. scalarization).
+  std::optional<InstructionCost> getMultipleResultIntrinsicVectorLibCallCost(
+      const IntrinsicCostAttributes &ICA, TTI::TargetCostKind CostKind,
+      RTLIB::Libcall LC, std::optional<unsigned> CallRetElementIndex = {}) {
+    Type *RetTy = ICA.getReturnType();
+    // Vector variants of the intrinsic can be mapped to a vector library call.
+    auto const *LibInfo = ICA.getLibInfo();
+    if (!LibInfo || !isa<StructType>(RetTy) ||
+        !isVectorizedStructTy(cast<StructType>(RetTy)))
+      return std::nullopt;
+
+    // Find associated libcall.
+    const char *LCName = getTLI()->getLibcallName(LC);
+    if (!LCName)
+      return std::nullopt;
+
+    // Search for a corresponding vector variant.
+    LLVMContext &Ctx = RetTy->getContext();
+    ElementCount VF = getVectorizedTypeVF(RetTy);
+    VecDesc const *VD = nullptr;
+    for (bool Masked : {false, true}) {
+      if ((VD = LibInfo->getVectorMappingInfo(LCName, VF, Masked)))
+        break;
+    }
+    if (!VD)
+      return std::nullopt;
+
+    // Cost the call + mask.
+    auto Cost =
+        thisT()->getCallInstrCost(nullptr, RetTy, ICA.getArgTypes(), CostKind);
+    if (VD->isMasked())
+      Cost += thisT()->getShuffleCost(
+          TargetTransformInfo::SK_Broadcast,
+          VectorType::get(IntegerType::getInt1Ty(Ctx), VF), {}, CostKind, 0,
+          nullptr, {});
+
+    // Lowering to a library call (with output pointers) may require us to emit
+    // reloads for the results.
+    for (auto [Idx, VectorTy] : enumerate(getContainedTypes(RetTy))) {
+      if (Idx == CallRetElementIndex)
+        continue;
+      Cost += thisT()->getMemoryOpCost(
+          Instruction::Load, VectorTy,
+          thisT()->getDataLayout().getABITypeAlign(VectorTy), 0, CostKind);
+    }
+    return Cost;
+  }
+  /* End downstream change: #87 */
+
 protected:
   explicit BasicTTIImplBase(const TargetMachine *TM, const DataLayout &DL)
       : BaseT(DL) {}
@@ -1716,9 +1777,11 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
 
     Type *RetTy = ICA.getReturnType();
 
-    ElementCount RetVF =
-        (RetTy->isVectorTy() ? cast<VectorType>(RetTy)->getElementCount()
-                             : ElementCount::getFixed(1));
+    /* Downstream change: #87 (sincos vectorization)*/
+    ElementCount RetVF = isVectorizedTy(RetTy) ? getVectorizedTypeVF(RetTy)
+                                               : ElementCount::getFixed(1);
+    /* End downstream change: #87 */
+
     const IntrinsicInst *I = ICA.getInst();
     const SmallVectorImpl<const Value *> &Args = ICA.getArgs();
     FastMathFlags FMF = ICA.getFlags();
@@ -1971,6 +2034,18 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
     }
     case Intrinsic::experimental_vector_match:
       return thisT()->getTypeBasedIntrinsicInstrCost(ICA, CostKind);
+      /* Downstream change: #87 (sincos vectorization)*/
+    case Intrinsic::sincos: {
+      Type *Ty = getContainedTypes(RetTy).front();
+      EVT VT = getTLI()->getValueType(DL, Ty);
+      RTLIB::Libcall LC = RTLIB::getFSINCOS(VT.getScalarType());
+      if (auto Cost =
+              getMultipleResultIntrinsicVectorLibCallCost(ICA, CostKind, LC))
+        return *Cost;
+      // Otherwise, fallback to default scalarization cost.
+      break;
+    }
+      /* End downstream change: #87 */
     }
 
     // Assume that we need to scalarize this intrinsic.)
@@ -1979,10 +2054,15 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
     InstructionCost ScalarizationCost = InstructionCost::getInvalid();
     if (RetVF.isVector() && !RetVF.isScalable()) {
       ScalarizationCost = 0;
-      if (!RetTy->isVoidTy())
-        ScalarizationCost += getScalarizationOverhead(
-            cast<VectorType>(RetTy),
-            /*Insert*/ true, /*Extract*/ false, CostKind);
+      /* Downstream change: #87 (sincos vectorization)*/
+      if (!RetTy->isVoidTy()) {
+        for (Type *VectorTy : getContainedTypes(RetTy)) {
+          ScalarizationCost += getScalarizationOverhead(
+              cast<VectorType>(VectorTy),
+              /*Insert=*/true, /*Extract=*/false, CostKind);
+        }
+      }
+      /* End downstream change: #87 */
       ScalarizationCost +=
           getOperandsScalarizationOverhead(Args, ICA.getArgTypes(), CostKind);
     }
@@ -2637,27 +2717,34 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
     // Else, assume that we need to scalarize this intrinsic. For math builtins
     // this will emit a costly libcall, adding call overhead and spills. Make it
     // very expensive.
-    if (auto *RetVTy = dyn_cast<VectorType>(RetTy)) {
+    /* Downstream change: #87 (sincos vectorization)*/
+    if (isVectorizedTy(RetTy)) {
+      ArrayRef<Type *> RetVTys = getContainedTypes(RetTy);
+
       // Scalable vectors cannot be scalarized, so return Invalid.
-      if (isa<ScalableVectorType>(RetTy) || any_of(Tys, [](const Type *Ty) {
-            return isa<ScalableVectorType>(Ty);
-          }))
+      if (any_of(concat<Type *const>(RetVTys, Tys),
+                 [](Type *Ty) { return isa<ScalableVectorType>(Ty); }))
         return InstructionCost::getInvalid();
 
-      InstructionCost ScalarizationCost =
-          SkipScalarizationCost
-              ? ScalarizationCostPassed
-              : getScalarizationOverhead(RetVTy, /*Insert*/ true,
-                                         /*Extract*/ false, CostKind);
+      InstructionCost ScalarizationCost = ScalarizationCostPassed;
+      if (!SkipScalarizationCost) {
+        ScalarizationCost = 0;
+        for (Type *RetVTy : RetVTys) {
+          ScalarizationCost += getScalarizationOverhead(
+              cast<VectorType>(RetVTy), /*Insert=*/true,
+              /*Extract=*/false, CostKind);
+        }
+      }
 
-      unsigned ScalarCalls = cast<FixedVectorType>(RetVTy)->getNumElements();
+      unsigned ScalarCalls = getVectorizedTypeVF(RetTy).getFixedValue();
       SmallVector<Type *, 4> ScalarTys;
       for (Type *Ty : Tys) {
         if (Ty->isVectorTy())
           Ty = Ty->getScalarType();
         ScalarTys.push_back(Ty);
       }
-      IntrinsicCostAttributes Attrs(IID, RetTy->getScalarType(), ScalarTys, FMF);
+      IntrinsicCostAttributes Attrs(IID, toScalarizedTy(RetTy), ScalarTys, FMF);
+      /* End downstream change: #87 */
       InstructionCost ScalarCost =
           thisT()->getIntrinsicInstrCost(Attrs, CostKind);
       for (Type *Ty : Tys) {

llvm/lib/Analysis/CostModel.cpp

@@ -17,13 +17,15 @@
 //===----------------------------------------------------------------------===//
 
 #include "llvm/Analysis/CostModel.h"
+#include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/PassManager.h"
 #include "llvm/Pass.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/raw_ostream.h"
+
 using namespace llvm;
 
 static cl::opt<TargetTransformInfo::TargetCostKind> CostKind(
@@ -42,25 +44,36 @@ static cl::opt<bool> TypeBasedIntrinsicCost("type-based-intrinsic-cost",
     cl::desc("Calculate intrinsics cost based only on argument types"),
     cl::init(false));
 
+/* Downstream change: #87 (sincos vectorization)*/
+static cl::opt<bool> PreferIntrinsicCost(
+    "prefer-intrinsic-cost",
+    cl::desc("Prefer using getIntrinsicInstrCost over getInstructionCost"),
+    cl::init(false));
+/* End downstream change: #87 */
+
 #define CM_NAME "cost-model"
 #define DEBUG_TYPE CM_NAME
 
 PreservedAnalyses CostModelPrinterPass::run(Function &F,
                                             FunctionAnalysisManager &AM) {
   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
+  // Downstream change: #87 (sincos vectorization)
+  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
   OS << "Printing analysis 'Cost Model Analysis' for function '" << F.getName() << "':\n";
   for (BasicBlock &B : F) {
     for (Instruction &Inst : B) {
       // TODO: Use a pass parameter instead of cl::opt CostKind to determine
       // which cost kind to print.
       InstructionCost Cost;
       auto *II = dyn_cast<IntrinsicInst>(&Inst);
-      if (II && TypeBasedIntrinsicCost) {
-        IntrinsicCostAttributes ICA(II->getIntrinsicID(), *II,
-                                    InstructionCost::getInvalid(), true);
+      /* Downstream change: #87 (sincos vectorization)*/
+      if (II && (PreferIntrinsicCost || TypeBasedIntrinsicCost)) {
+        IntrinsicCostAttributes ICA(
+            II->getIntrinsicID(), *II, InstructionCost::getInvalid(),
+            /*TypeBasedOnly=*/TypeBasedIntrinsicCost, &TLI);
         Cost = TTI.getIntrinsicInstrCost(ICA, CostKind);
-      }
-      else {
+      } else {
+        /* End downstream change: #87 */
         Cost = TTI.getInstructionCost(&Inst, CostKind);
       }
 

llvm/lib/Analysis/TargetTransformInfo.cpp

@@ -67,11 +67,13 @@ bool HardwareLoopInfo::canAnalyze(LoopInfo &LI) {
   return true;
 }
 
+/* Downstream change: #87 (sincos vectorization)*/
 IntrinsicCostAttributes::IntrinsicCostAttributes(
     Intrinsic::ID Id, const CallBase &CI, InstructionCost ScalarizationCost,
-    bool TypeBasedOnly)
+    bool TypeBasedOnly, const TargetLibraryInfo *LibInfo)
     : II(dyn_cast<IntrinsicInst>(&CI)), RetTy(CI.getType()), IID(Id),
-      ScalarizationCost(ScalarizationCost) {
+      ScalarizationCost(ScalarizationCost), LibInfo(LibInfo) {
+  /* End downstream change: #87 */
 
   if (const auto *FPMO = dyn_cast<FPMathOperator>(&CI))
     FMF = FPMO->getFastMathFlags();
@@ -101,13 +103,14 @@ IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *Ty,
     ParamTys.push_back(Argument->getType());
 }
 
-IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *RTy,
-                                                 ArrayRef<const Value *> Args,
-                                                 ArrayRef<Type *> Tys,
-                                                 FastMathFlags Flags,
-                                                 const IntrinsicInst *I,
-                                                 InstructionCost ScalarCost)
-    : II(I), RetTy(RTy), IID(Id), FMF(Flags), ScalarizationCost(ScalarCost) {
+/* Downstream change: #87 (sincos vectorization)*/
+IntrinsicCostAttributes::IntrinsicCostAttributes(
+    Intrinsic::ID Id, Type *RTy, ArrayRef<const Value *> Args,
+    ArrayRef<Type *> Tys, FastMathFlags Flags, const IntrinsicInst *I,
+    InstructionCost ScalarCost, TargetLibraryInfo const *LibInfo)
+    : II(I), RetTy(RTy), IID(Id), FMF(Flags), ScalarizationCost(ScalarCost),
+      LibInfo(LibInfo) {
+  /* End downstream change: #87 */
   ParamTys.insert(ParamTys.begin(), Tys.begin(), Tys.end());
   Arguments.insert(Arguments.begin(), Args.begin(), Args.end());
 }

llvm/lib/Analysis/VectorUtils.cpp

@@ -72,6 +72,8 @@ bool llvm::isTriviallyVectorizable(Intrinsic::ID ID) {
   case Intrinsic::atan2:
   case Intrinsic::sin:
   case Intrinsic::cos:
+  // Downstream change: #87 (sincos vectorization)
+  case Intrinsic::sincos:
   case Intrinsic::tan:
   case Intrinsic::sinh:
   case Intrinsic::cosh:
@@ -179,6 +181,8 @@ bool llvm::isVectorIntrinsicWithOverloadTypeAtArg(
   case Intrinsic::ucmp:
   case Intrinsic::scmp:
     return OpdIdx == -1 || OpdIdx == 0;
+    // Downstream change: #87 (sincos vectorization)
+  case Intrinsic::sincos:
   case Intrinsic::is_fpclass:
   case Intrinsic::vp_is_fpclass:
     return OpdIdx == 0;

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -2884,8 +2884,11 @@ LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI,
                  std::back_inserter(ParamTys),
                  [&](Type *Ty) { return maybeVectorizeType(Ty, VF); });
 
+  /* Downstream change: #87 (sincos vectorization)*/
   IntrinsicCostAttributes CostAttrs(ID, RetTy, Arguments, ParamTys, FMF,
-                                    dyn_cast<IntrinsicInst>(CI));
+                                    dyn_cast<IntrinsicInst>(CI),
+                                    InstructionCost::getInvalid(), TLI);
+  /* End downstream change: #87 */
   return TTI.getIntrinsicInstrCost(CostAttrs, CostKind);
 }
 

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -1149,9 +1149,12 @@ InstructionCost VPWidenIntrinsicRecipe::computeCost(ElementCount VF,
 
   // TODO: Rework TTI interface to avoid reliance on underlying IntrinsicInst.
   FastMathFlags FMF = hasFastMathFlags() ? getFastMathFlags() : FastMathFlags();
+  /* Downstream change: #87 (sincos vectorization)*/
   IntrinsicCostAttributes CostAttrs(
       VectorIntrinsicID, RetTy, Arguments, ParamTys, FMF,
-      dyn_cast_or_null<IntrinsicInst>(getUnderlyingValue()));
+      dyn_cast_or_null<IntrinsicInst>(getUnderlyingValue()),
+      InstructionCost::getInvalid(), &Ctx.TLI);
+  /* End downstream change: #87 */
   return Ctx.TTI.getIntrinsicInstrCost(CostAttrs, Ctx.CostKind);
 }