LoopVectorize: We don't need an identity element for min/max reductions

We can just use the initial element that feeds the reduction.

  max(max(x, y), z) == max(max(x,y), max(x,z))

radar://13723044

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181141 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -498,8 +498,7 @@ public:
 
   /// This function returns the identity element (or neutral element) for
   /// the operation K.
-  static Constant *getReductionIdentity(ReductionKind K, Type *Tp,
+  static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
-                                        MinMaxReductionKind MinMaxK);
 private:
   /// Check if a single basic block loop is vectorizable.
   /// At this point we know that this is a loop with a constant trip count
@@ -1501,8 +1500,7 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
 /// This function returns the identity element (or neutral element) for
 /// the operation K.
 Constant*
-LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp,
+LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
-                                                MinMaxReductionKind MinMaxK) {
   switch (K) {
   case RK_IntegerXor:
   case RK_IntegerAdd:
@@ -1521,24 +1519,6 @@ LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp,
   case  RK_FloatAdd:
     // Adding zero to a number does not change it.
     return ConstantFP::get(Tp, 0.0L);
-  case  RK_IntegerMinMax:
-    switch(MinMaxK) {
-    default: llvm_unreachable("Unknown min/max predicate");
-    case MRK_UIntMin:
-      return ConstantInt::getAllOnesValue(Tp);
-    case MRK_UIntMax:
-      return ConstantInt::get(Tp, 0);
-    case MRK_SIntMin: {
-      unsigned BitWidth = Tp->getPrimitiveSizeInBits();
-      return ConstantInt::get(Tp->getContext(),
-                              APInt::getSignedMaxValue(BitWidth));
-    }
-    case LoopVectorizationLegality::MRK_SIntMax: {
-      unsigned BitWidth = Tp->getPrimitiveSizeInBits();
-      return ConstantInt::get(Tp->getContext(),
-                              APInt::getSignedMinValue(BitWidth));
-    }
-    }
   default:
     llvm_unreachable("Unknown reduction kind");
   }
@@ -1761,16 +1741,23 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
 
     // Find the reduction identity variable. Zero for addition, or, xor,
     // one for multiplication, -1 for And.
+    Value *Identity;
+    Value *VectorStart;
+    if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax)
+      // MinMax reduction have the start value as their identify.
+      VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
+                                                         "minmax.ident");
+    else {
       Constant *Iden =
         LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
-                                                      VecTy->getScalarType(),
+                                                        VecTy->getScalarType());
-                                                      RdxDesc.MinMaxKind);
+      Identity = ConstantVector::getSplat(VF, Iden);
-    Constant *Identity = ConstantVector::getSplat(VF, Iden);
 
       // This vector is the Identity vector where the first element is the
       // incoming scalar reduction.
-    Value *VectorStart = Builder.CreateInsertElement(Identity,
+      VectorStart = Builder.CreateInsertElement(Identity,
                                                 RdxDesc.StartValue, Zero);
+    }
 
     // Fix the vector-loop phi.
     // We created the induction variable so we know that the

test/Transforms/LoopVectorize/minmax_reduction.ll

@@ -6,8 +6,11 @@ target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f3
 
 ; Signed tests.
 
-; Turn this into a max reduction.
+; Turn this into a max reduction. Make sure we use a splat to initialize the
+; vector for the reduction.
 ; CHECK: @max_red
+; CHECK: %[[VAR:.*]] = insertelement <2 x i32> undef, i32 %max, i32 0
+; CHECK: {{.*}} = shufflevector <2 x i32> %[[VAR]], <2 x i32> undef, <2 x i32> zeroinitializer
 ; CHECK: icmp sgt <2 x i32>
 ; CHECK: select <2 x i1>
 ; CHECK: middle.block