[LV] Teach the vectorizer to cost and vectorize llvm.sincos intrinsics

This teaches the loop vectorizer that `llvm.sincos` is trivially
vectorizable. Additionally, this patch updates the cost model to cost
intrinsics that return multiple values correctly. Previously, the cost
model only thought intrinsics that return `VectorType` need scalarizing,
which meant it cost intrinsics that return multiple vectors (that need
scalarizing) way too cheap (giving it the cost of a single function
call).

The `llvm.sincos` intrinsic also has a custom cost when a vector
function library is available, as certain VFs can be expanded (later in
code-gen) to a vector function, reducing the cost to a single call (+
the possible loads from the vector function returns values via output
pointers).

Author: MacDue

llvm/include/llvm/Analysis/TargetTransformInfo.h

@@ -126,6 +126,7 @@ 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();
+  TargetLibraryInfo const *LibInfo = nullptr;
 
 public:
   IntrinsicCostAttributes(
@@ -145,7 +146,8 @@ class 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);
 
   Intrinsic::ID getID() const { return IID; }
   const IntrinsicInst *getInst() const { return II; }
@@ -154,6 +156,7 @@ class IntrinsicCostAttributes {
   InstructionCost getScalarizationCost() const { return ScalarizationCost; }
   const SmallVectorImpl<const Value *> &getArgs() const { return Arguments; }
   const SmallVectorImpl<Type *> &getArgTypes() const { return ParamTys; }
+  const TargetLibraryInfo *getLibInfo() const { return LibInfo; }
 
   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"
@@ -1718,8 +1719,8 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
     Type *RetTy = ICA.getReturnType();
 
     ElementCount RetVF =
-        (RetTy->isVectorTy() ? cast<VectorType>(RetTy)->getElementCount()
-                             : ElementCount::getFixed(1));
+        isVectorizedTy(RetTy) ? getVectorizedTypeVF(RetTy) : ElementCount::getFixed(1);
+
     const IntrinsicInst *I = ICA.getInst();
     const SmallVectorImpl<const Value *> &Args = ICA.getArgs();
     FastMathFlags FMF = ICA.getFlags();
@@ -1972,6 +1973,47 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
     }
     case Intrinsic::experimental_vector_match:
       return thisT()->getTypeBasedIntrinsicInstrCost(ICA, CostKind);
+    case Intrinsic::sincos: {
+      // Vector variants of llvm.sincos can be mapped to a vector library call.
+      auto const *LibInfo = ICA.getLibInfo();
+      if (!LibInfo || !isVectorizedTy(RetTy))
+        break;
+
+      // Find associated libcall.
+      VectorType *VectorTy = cast<VectorType>(getContainedTypes(RetTy).front());
+      EVT VT = getTLI()->getValueType(DL, VectorTy);
+      RTLIB::Libcall LC = RTLIB::getFSINCOS(VT.getVectorElementType());
+      const char *LCName = getTLI()->getLibcallName(LC);
+      if (!LC || !LCName)
+        break;
+
+      // Search for a corresponding vector variant.
+      LLVMContext &Ctx = RetTy->getContext();
+      auto VF = getVectorizedTypeVF(RetTy);
+      VecDesc const *VD = nullptr;
+      for (bool Masked : {false, true}) {
+        if ((VD = LibInfo->getVectorMappingInfo(LCName, VF, Masked)))
+          break;
+      }
+      if (!VD)
+        break;
+
+      // 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 sincos library call (with output pointers) may require us
+      // to emit reloads for the results.
+      Cost +=
+          thisT()->getMemoryOpCost(Instruction::Load, VectorTy,
+                          thisT()->getDataLayout().getABITypeAlign(VectorTy), 0,
+                          CostKind) *
+          2;
+      return Cost;
+    }
     }
 
     // Assume that we need to scalarize this intrinsic.)
@@ -1980,10 +2022,13 @@ 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);
+      if (!RetTy->isVoidTy()) {
+        for (Type *VectorTy : getContainedTypes(RetTy)) {
+          ScalarizationCost += getScalarizationOverhead(
+              cast<VectorType>(VectorTy),
+              /*Insert*/ true, /*Extract*/ false, CostKind);
+        }
+      }
       ScalarizationCost +=
           getOperandsScalarizationOverhead(Args, ICA.getArgTypes(), CostKind);
     }
@@ -2609,27 +2654,32 @@ 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)) {
+    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);
       InstructionCost ScalarCost =
           thisT()->getIntrinsicInstrCost(Attrs, CostKind);
       for (Type *Ty : Tys) {

llvm/lib/Analysis/TargetTransformInfo.cpp

@@ -101,13 +101,12 @@ 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) {
+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) {
   ParamTys.insert(ParamTys.begin(), Tys.begin(), Tys.end());
   Arguments.insert(Arguments.begin(), Args.begin(), Args.end());
 }

llvm/lib/Analysis/VectorUtils.cpp

@@ -72,6 +72,7 @@ bool llvm::isTriviallyVectorizable(Intrinsic::ID ID) {
   case Intrinsic::atan2:
   case Intrinsic::sin:
   case Intrinsic::cos:
+  case Intrinsic::sincos:
   case Intrinsic::tan:
   case Intrinsic::sinh:
   case Intrinsic::cosh:
@@ -179,6 +180,7 @@ bool llvm::isVectorIntrinsicWithOverloadTypeAtArg(
   case Intrinsic::ucmp:
   case Intrinsic::scmp:
     return OpdIdx == -1 || OpdIdx == 0;
+  case Intrinsic::sincos:
   case Intrinsic::is_fpclass:
   case Intrinsic::vp_is_fpclass:
     return OpdIdx == 0;

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -2971,7 +2971,8 @@ LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI,
                  [&](Type *Ty) { return maybeVectorizeType(Ty, VF); });
 
   IntrinsicCostAttributes CostAttrs(ID, RetTy, Arguments, ParamTys, FMF,
-                                    dyn_cast<IntrinsicInst>(CI));
+                                    dyn_cast<IntrinsicInst>(CI),
+                                    InstructionCost::getInvalid(), TLI);
   return TTI.getIntrinsicInstrCost(CostAttrs,
                                    TargetTransformInfo::TCK_RecipThroughput);
 }

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -1112,7 +1112,8 @@ InstructionCost VPWidenIntrinsicRecipe::computeCost(ElementCount VF,
   FastMathFlags FMF = hasFastMathFlags() ? getFastMathFlags() : FastMathFlags();
   IntrinsicCostAttributes CostAttrs(
       VectorIntrinsicID, RetTy, Arguments, ParamTys, FMF,
-      dyn_cast_or_null<IntrinsicInst>(getUnderlyingValue()));
+      dyn_cast_or_null<IntrinsicInst>(getUnderlyingValue()),
+      InstructionCost::getInvalid(), &Ctx.TLI);
   return Ctx.TTI.getIntrinsicInstrCost(CostAttrs, CostKind);
 }