[llvm] [SLP] Inefficient cost-modelling and codegen for reductions with slp-… (PR #197875)

[Sushant Gokhale via llvm-commits]

================
@@ -29189,34 +29190,48 @@ class HorizontalReduction {
 private:
   /// Creates the reduction from the given \p Vec vector value with the given
   /// scale \p Scale and signedness \p IsSigned.
-  Value *createSingleOp(IRBuilderBase &Builder, const TargetTransformInfo &TTI,
-                        Value *Vec, unsigned Scale, bool IsSigned, Type *DestTy,
+  Value *createSingleOp(BoUpSLP &V, IRBuilderBase &Builder,
+                        const TargetTransformInfo &TTI, Value *Vec,
+                        unsigned Scale, bool IsSigned, Type *DestTy,
                         bool ReducedInTree) {
     Value *Rdx;
     if (ReducedInTree) {
       Rdx = Vec;
     } else if (auto *VecTy = dyn_cast<FixedVectorType>(DestTy)) {
       unsigned DestTyNumElements = getNumElements(VecTy);
       unsigned VF = getNumElements(Vec->getType()) / DestTyNumElements;
-      Rdx = PoisonValue::get(
-          getWidenedType(Vec->getType()->getScalarType(), DestTyNumElements));
-      for (unsigned I : seq<unsigned>(DestTyNumElements)) {
-        // Do reduction for each lane.
-        // e.g., do reduce add for
-        // VL[0] = <4 x Ty> <a, b, c, d>
-        // VL[1] = <4 x Ty> <e, f, g, h>
-        // Lane[0] = <2 x Ty> <a, e>
-        // Lane[1] = <2 x Ty> <b, f>
-        // Lane[2] = <2 x Ty> <c, g>
-        // Lane[3] = <2 x Ty> <d, h>
-        // result[0] = reduce add Lane[0]
-        // result[1] = reduce add Lane[1]
-        // result[2] = reduce add Lane[2]
-        // result[3] = reduce add Lane[3]
-        SmallVector<int, 16> Mask = createStrideMask(I, DestTyNumElements, VF);
-        Value *Lane = Builder.CreateShuffleVector(Vec, Mask);
-        Rdx = Builder.CreateInsertElement(
-            Rdx, emitReduction(Lane, Builder, &TTI, DestTy), I);
+      unsigned RdxOpcode = RecurrenceDescriptor::getOpcode(RdxKind);
+      Rdx = nullptr;
+      /*
+        e.g. Consider vector reduce add.
+
+        Initial reduction is
+        %add0 = add <4 x i32> zeroinitializer, %A
+        %add1 = add <4 x i32> %add0, %B
+
+        where,
+        %A = <a, b, c, d>
+        %B = <e, f, g, h>
+        %add1 = A + B = <a + e, b + f, c + g, d + h>
+
+        After revectorization with VF=2 and Vec = <a, b, c, d, e, f, g, h>,
+        the reduction can be expressed as:
+        %A = shufflevector <8 x i32> %Vec, <8 x i32> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+        %B = shufflevector <8 x i32> %Vec, <8 x i32> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+        %add0 = add <4 x i32> zeroinitializer, %A
+        %add1 = add <4 x i32> %add0, %B
+      */
----------------
sushgokh wrote:

done

https://github.com/llvm/llvm-project/pull/197875