[Vectorize] Fix vectorization, scalarization and folding of llvm.is.fpclass

[foad]

Summary

llvm.is.fpclass is different from other vectorizable intrinsics in that
it is overloaded on an argument type, not on the return type.

[foad]

Fold D148803 into this one.

[fhahn]

Similar to D132344, but it look like this patch has stalled.

llvm/lib/Analysis/VectorUtils.cpp

131

I'm not sure about the change in the default here, IICU this is now considering any other intrinsic to have the return type overloaded?

llvm/test/Transforms/LoopVectorize/ARM/is_fpclass.ll

5

The triple doesn't match the sub-directory, aarch64- needs to be in the AArch64 subdirectory.

But does the test need to be target specific in the first place? Could you just use -force-vector-width=4 -force-vector-interleave=1 or something, without a triple?

Also, would be good to add the test separately.

[foad]

llvm/lib/Analysis/VectorUtils.cpp

131

Yes. All callers of isVectorIntrinsicWithOverloadTypeAtArg were already making this assumption implicitly. I have changed them to make an explicit check for isVectorIntrinsicWithOverloadTypeAtArg(ID, -1).

llvm/test/Transforms/LoopVectorize/ARM/is_fpclass.ll

5

Thanks. I didn't know how to remove the triple. I will try the -force-vector-* options.

[foad]

Rebase on target-independent loop vectorize test.

[dmgreen]

Are there any test for the Scalarizer and SLP vectorizer yet? Similar to D124358.

[foad]

Add tests for scalarizer and SLP vectorizer.

[foad]

In D148905#4291721, @dmgreen wrote:

Are there any test for the Scalarizer and SLP vectorizer yet? Similar to D124358.

Added them, thanks!

[foad]

Rebase.

[fhahn]

LGTM, thanks!

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

485

Could you add a comment here with an explanation? I.e. we are adding the return type if the intrinsic is overloaded on it.

Might be good to add this in all updated places in the patch.

llvm/include/llvm/Analysis/VectorUtils.h

 /// for the scalar form of the intrinsic and all vectors (or scalars handled by
 /// isVectorIntrinsicWithScalarOpAtArg) for the vector form of the intrinsic.
 bool isTriviallyVectorizable(Intrinsic::ID ID);
 
 /// Identifies if the vector form of the intrinsic has a scalar operand.
 bool isVectorIntrinsicWithScalarOpAtArg(Intrinsic::ID ID,
                                         unsigned ScalarOpdIdx);
 
-/// Identifies if the vector form of the intrinsic has a operand that has
-/// an overloaded type.
-bool isVectorIntrinsicWithOverloadTypeAtArg(Intrinsic::ID ID, unsigned OpdIdx);
+/// Identifies if the vector form of the intrinsic is overloaded on the type of
+/// the operand at index \p OpdIdx, or on the return type if \p OpdIdx is -1.
+bool isVectorIntrinsicWithOverloadTypeAtArg(Intrinsic::ID ID, int OpdIdx);
 
 /// Returns intrinsic ID for call.
 /// For the input call instruction it finds mapping intrinsic and returns
 /// its intrinsic ID, in case it does not found it return not_intrinsic.
 Intrinsic::ID getVectorIntrinsicIDForCall(const CallInst *CI,
                                           const TargetLibraryInfo *TLI);
 
 /// Given a vector and an element number, see if the scalar value is

llvm/lib/Analysis/VectorUtils.cpp

bool llvm::isTriviallyVectorizable(Intrinsic::ID ID) {
   case Intrinsic::trunc:
   case Intrinsic::rint:
   case Intrinsic::nearbyint:
   case Intrinsic::round:
   case Intrinsic::roundeven:
   case Intrinsic::pow:
   case Intrinsic::fma:
   case Intrinsic::fmuladd:
+  case Intrinsic::is_fpclass:
   case Intrinsic::powi:
   case Intrinsic::canonicalize:
   case Intrinsic::fptosi_sat:
   case Intrinsic::fptoui_sat:
     return true;
   default:
     return false;
   }
 }
 
 /// Identifies if the vector form of the intrinsic has a scalar operand.
 bool llvm::isVectorIntrinsicWithScalarOpAtArg(Intrinsic::ID ID,
                                               unsigned ScalarOpdIdx) {
   switch (ID) {
   case Intrinsic::abs:
   case Intrinsic::ctlz:
   case Intrinsic::cttz:
+  case Intrinsic::is_fpclass:
   case Intrinsic::powi:
     return (ScalarOpdIdx == 1);
   case Intrinsic::smul_fix:
   case Intrinsic::smul_fix_sat:
   case Intrinsic::umul_fix:
   case Intrinsic::umul_fix_sat:
     return (ScalarOpdIdx == 2);
   default:
     return false;
   }
 }
 
 bool llvm::isVectorIntrinsicWithOverloadTypeAtArg(Intrinsic::ID ID,
-                                                  unsigned OpdIdx) {
+                                                  int OpdIdx) {
   switch (ID) {
   case Intrinsic::fptosi_sat:
   case Intrinsic::fptoui_sat:
+    return OpdIdx == -1 || OpdIdx == 0;
+  case Intrinsic::is_fpclass:
     return OpdIdx == 0;
   case Intrinsic::powi:
-    return OpdIdx == 1;
+    return OpdIdx == -1 || OpdIdx == 1;
   default:
-    return false;
+    return OpdIdx == -1;
   }
 }
 
 /// Returns intrinsic ID for call.
 /// For the input call instruction it finds mapping intrinsic and returns
 /// its ID, in case it does not found it return not_intrinsic.
 Intrinsic::ID llvm::getVectorIntrinsicIDForCall(const CallInst *CI,
                                                 const TargetLibraryInfo *TLI) {

llvm/lib/Transforms/Scalar/Scalarizer.cpp

bool ScalarizerVisitor::splitCall(CallInst &CI) {
   unsigned NumArgs = CI.arg_size();
 
   ValueVector ScalarOperands(NumArgs);
   SmallVector<Scatterer, 8> Scattered(NumArgs);
 
   Scattered.resize(NumArgs);
 
   SmallVector<llvm::Type *, 3> Tys;
+  if (isVectorIntrinsicWithOverloadTypeAtArg(ID, -1))
     Tys.push_back(VT->getScalarType());
 
   // Assumes that any vector type has the same number of elements as the return
   // vector type, which is true for all current intrinsics.
   for (unsigned I = 0; I != NumArgs; ++I) {
     Value *OpI = CI.getOperand(I);
     if (OpI->getType()->isVectorTy()) {
       Scattered[I] = scatter(&CI, OpI);
       assert(Scattered[I].size() == NumElems && "mismatched call operands");

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

case Instruction::Call: {
       Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
 
       auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI);
       bool UseIntrinsic = ID != Intrinsic::not_intrinsic &&
                           VecCallCosts.first <= VecCallCosts.second;
 
       Value *ScalarArg = nullptr;
       std::vector<Value *> OpVecs;
-      SmallVector<Type *, 2> TysForDecl =
-          {FixedVectorType::get(CI->getType(), E->Scalars.size())};
+      SmallVector<Type *, 2> TysForDecl;
+      if (isVectorIntrinsicWithOverloadTypeAtArg(IID, -1))
+        TysForDecl.push_back(
+            FixedVectorType::get(CI->getType(), E->Scalars.size()));
       for (int j = 0, e = CI->arg_size(); j < e; ++j) {
         ValueList OpVL;
         // Some intrinsics have scalar arguments. This argument should not be
         // vectorized.
         if (UseIntrinsic && isVectorIntrinsicWithScalarOpAtArg(IID, j)) {
           CallInst *CEI = cast<CallInst>(VL0);
           ScalarArg = CEI->getArgOperand(j);
           OpVecs.push_back(CEI->getArgOperand(j));

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

 
 void VPWidenCallRecipe::execute(VPTransformState &State) {
   auto &CI = *cast<CallInst>(getUnderlyingInstr());
   assert(!isa<DbgInfoIntrinsic>(CI) &&
          "DbgInfoIntrinsic should have been dropped during VPlan construction");
   State.setDebugLocFromInst(&CI);
 
   for (unsigned Part = 0; Part < State.UF; ++Part) {
-    SmallVector<Type *, 2> TysForDecl = {CI.getType()};
+    SmallVector<Type *, 2> TysForDecl;
+    if (isVectorIntrinsicWithOverloadTypeAtArg(VectorIntrinsicID, -1)) {
+      TysForDecl.push_back(
+          State.VF.isVector()
+              ? VectorType::get(CI.getType()->getScalarType(), State.VF)
+              : CI.getType());
+    }
     SmallVector<Value *, 4> Args;
     for (const auto &I : enumerate(operands())) {
       // Some intrinsics have a scalar argument - don't replace it with a
       // vector.
       Value *Arg;
       if (VectorIntrinsicID == Intrinsic::not_intrinsic ||
           !isVectorIntrinsicWithScalarOpAtArg(VectorIntrinsicID, I.index()))
         Arg = State.get(I.value(), Part);
       else
         Arg = State.get(I.value(), VPIteration(0, 0));
       if (isVectorIntrinsicWithOverloadTypeAtArg(VectorIntrinsicID, I.index()))
         TysForDecl.push_back(Arg->getType());
       Args.push_back(Arg);
     }
 
     Function *VectorF;
     if (VectorIntrinsicID != Intrinsic::not_intrinsic) {
       // Use vector version of the intrinsic.
-      if (State.VF.isVector())
-        TysForDecl[0] =
-            VectorType::get(CI.getType()->getScalarType(), State.VF);
       Module *M = State.Builder.GetInsertBlock()->getModule();
       VectorF = Intrinsic::getDeclaration(M, VectorIntrinsicID, TysForDecl);
       assert(VectorF && "Can't retrieve vector intrinsic.");
     } else {
 #ifndef NDEBUG
       assert(Variant != nullptr && "Can't create vector function.");
 #endif
       VectorF = Variant;

llvm/test/Transforms/InstSimplify/is_fpclass.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
+; RUN: opt < %s -S -passes=instsimplify | FileCheck %s
+
+define <2 x i1> @f() {
+; CHECK-LABEL: define <2 x i1> @f() {
+; CHECK-NEXT:    ret <2 x i1> zeroinitializer
+;
+  %i = call <2 x i1> @llvm.is.fpclass.v2f16(<2 x half> <half 0xH7C00, half 0xH7C00>, i32 3)
+  ret <2 x i1> %i
+}
+
+declare <2 x i1> @llvm.is.fpclass.v2f16(<2 x half>, i32 immarg)

llvm/test/Transforms/LoopVectorize/ARM/is_fpclass.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
+; RUN: opt < %s -passes=loop-vectorize -S | FileCheck %s
+
+target triple = "aarch64-unknown-linux-gnu"
+
+define void @d() {
+; CHECK-LABEL: define void @d() {
+; CHECK-NEXT:  bb:
+; CHECK-NEXT:    br i1 true, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK:       vector.ph:
+; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 2
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr float, ptr @d, i64 [[TMP0]]
+; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr float, ptr @d, i64 [[TMP1]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call <2 x i1> @llvm.is.fpclass.v2f32(<2 x float> zeroinitializer, i32 0)
+; CHECK-NEXT:    [[TMP5:%.*]] = call <2 x i1> @llvm.is.fpclass.v2f32(<2 x float> zeroinitializer, i32 0)
+; CHECK-NEXT:    [[TMP6:%.*]] = select <2 x i1> [[TMP4]], <2 x float> zeroinitializer, <2 x float> zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = select <2 x i1> [[TMP5]], <2 x float> zeroinitializer, <2 x float> zeroinitializer
+; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr float, ptr [[TMP2]], i32 0
+; CHECK-NEXT:    store <2 x float> [[TMP6]], ptr [[TMP8]], align 4
+; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr float, ptr [[TMP2]], i32 2
+; CHECK-NEXT:    store <2 x float> [[TMP7]], ptr [[TMP9]], align 4
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 0
+; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK:       middle.block:
+; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 0, 0
+; CHECK-NEXT:    br i1 [[CMP_N]], label [[BB1:%.*]], label [[SCALAR_PH]]
+; CHECK:       scalar.ph:
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 0, [[MIDDLE_BLOCK]] ], [ 0, [[BB:%.*]] ]
+; CHECK-NEXT:    br label [[BB2:%.*]]
+; CHECK:       bb1:
+; CHECK-NEXT:    ret void
+; CHECK:       bb2:
+; CHECK-NEXT:    [[I:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[I7:%.*]], [[BB2]] ]
+; CHECK-NEXT:    [[I3:%.*]] = load float, ptr null, align 4
+; CHECK-NEXT:    [[I4:%.*]] = getelementptr float, ptr @d, i64 [[I]]
+; CHECK-NEXT:    [[I5:%.*]] = tail call i1 @llvm.is.fpclass.f32(float 0.000000e+00, i32 0)
+; CHECK-NEXT:    [[I6:%.*]] = select i1 [[I5]], float 0.000000e+00, float 0.000000e+00
+; CHECK-NEXT:    store float [[I6]], ptr [[I4]], align 4
+; CHECK-NEXT:    [[I7]] = add i64 [[I]], 1
+; CHECK-NEXT:    [[I8:%.*]] = icmp eq i64 [[I7]], 0
+; CHECK-NEXT:    br i1 [[I8]], label [[BB1]], label [[BB2]], !llvm.loop [[LOOP3:![0-9]+]]
+;
+bb:
+  br label %bb2
+
+bb1:
+  ret void
+
+bb2:
+  %i = phi i64 [ 0, %bb ], [ %i7, %bb2 ]
+  %i3 = load float, ptr null, align 4
+  %i4 = getelementptr float, ptr @d, i64 %i
+  %i5 = tail call i1 @llvm.is.fpclass.f32(float 0.0, i32 0)
+  %i6 = select i1 %i5, float 0.0, float 0.0
+  store float %i6, ptr %i4, align 4
+  %i7 = add i64 %i, 1
+  %i8 = icmp eq i64 %i7, 0
+  br i1 %i8, label %bb1, label %bb2
+}
+
+declare i1 @llvm.is.fpclass.f32(float, i32 immarg)