[SLPVectorizer] Widen constant strided loads. (#162324)

Given a set of pointers, check if they can be rearranged as follows (%s is a constant):
%b + 0 * %s + 0
%b + 0 * %s + 1
%b + 0 * %s + 2
...
%b + 0 * %s + w

%b + 1 * %s + 0
%b + 1 * %s + 1
%b + 1 * %s + 2
...
%b + 1 * %s + w
...

If the pointers can be rearanged in the above pattern, it means that the
memory can be accessed with a strided loads of width `w` and stride `%s`.

Author: mgudim

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -2245,7 +2245,6 @@ class slpvectorizer::BoUpSLP {
   /// Return true if an array of scalar loads can be replaced with a strided
   ///  load (with constant stride).
   ///
-  ///  TODO:
   ///  It is possible that the load gets "widened". Suppose that originally each
   ///  load loads `k` bytes and `PointerOps` can be arranged as follows (`%s` is
   ///  constant): %b + 0 * %s + 0 %b + 0 * %s + 1 %b + 0 * %s + 2
@@ -7008,32 +7007,93 @@ bool BoUpSLP::analyzeConstantStrideCandidate(
     const SmallVectorImpl<unsigned> &SortedIndices, const int64_t Diff,
     Value *Ptr0, Value *PtrN, StridedPtrInfo &SPtrInfo) const {
   const size_t Sz = PointerOps.size();
-  if (!isStridedLoad(PointerOps, ScalarTy, Alignment, Diff, Sz))
+  SmallVector<int64_t> SortedOffsetsFromBase(Sz);
+  // Go through `PointerOps` in sorted order and record offsets from `Ptr0`.
+  for (unsigned I : seq<unsigned>(Sz)) {
+    Value *Ptr =
+        SortedIndices.empty() ? PointerOps[I] : PointerOps[SortedIndices[I]];
+    SortedOffsetsFromBase[I] =
+        *getPointersDiff(ScalarTy, Ptr0, ScalarTy, Ptr, *DL, *SE);
+  }
+
+  // The code below checks that `SortedOffsetsFromBase` looks as follows:
+  // ```
+  // [
+  //   (e_{0, 0}, e_{0, 1}, ..., e_{0, GroupSize - 1}), // first group
+  //   (e_{1, 0}, e_{1, 1}, ..., e_{1, GroupSize - 1}), // secon group
+  //   ...
+  //   (e_{NumGroups - 1, 0}, e_{NumGroups - 1, 1}, ..., e_{NumGroups - 1,
+  //   GroupSize - 1}), // last group
+  // ]
+  // ```
+  // The distance between consecutive elements within each group should all be
+  // the same `StrideWithinGroup`. The distance between the first elements of
+  // consecutive groups should all be the same `StrideBetweenGroups`.
+
+  int64_t StrideWithinGroup =
+      SortedOffsetsFromBase[1] - SortedOffsetsFromBase[0];
+  // Determine size of the first group. Later we will check that all other
+  // groups have the same size.
+  auto IsEndOfGroupIndex = [=, &SortedOffsetsFromBase](unsigned Idx) {
+    return SortedOffsetsFromBase[Idx] - SortedOffsetsFromBase[Idx - 1] !=
+           StrideWithinGroup;
+  };
+  auto Indices = seq<unsigned>(1, Sz);
+  auto FoundIt = llvm::find_if(Indices, IsEndOfGroupIndex);
+  unsigned GroupSize = FoundIt != Indices.end() ? *FoundIt : Sz;
+
+  unsigned VecSz = Sz;
+  Type *NewScalarTy = ScalarTy;
+
+  // Quick detour: at this point we can say what the type of strided load would
+  // be if all the checks pass. Check if this type is legal for the target.
+  bool NeedsWidening = Sz != GroupSize;
+  if (NeedsWidening) {
+    if (Sz % GroupSize != 0)
+      return false;
+
+    if (StrideWithinGroup != 1)
+      return false;
+    VecSz = Sz / GroupSize;
+    NewScalarTy = Type::getIntNTy(
+        SE->getContext(),
+        DL->getTypeSizeInBits(ScalarTy).getFixedValue() * GroupSize);
+  }
+
+  if (!isStridedLoad(PointerOps, NewScalarTy, Alignment, Diff, VecSz))
     return false;
 
-  int64_t Stride = Diff / static_cast<int64_t>(Sz - 1);
+  int64_t StrideIntVal = StrideWithinGroup;
+  if (NeedsWidening) {
+    // Continue with checking the "shape" of `SortedOffsetsFromBase`.
+    // Check that the strides between groups are all the same.
+    unsigned CurrentGroupStartIdx = GroupSize;
+    int64_t StrideBetweenGroups =
+        SortedOffsetsFromBase[GroupSize] - SortedOffsetsFromBase[0];
+    StrideIntVal = StrideBetweenGroups;
+    for (; CurrentGroupStartIdx < Sz; CurrentGroupStartIdx += GroupSize) {
+      if (SortedOffsetsFromBase[CurrentGroupStartIdx] -
+              SortedOffsetsFromBase[CurrentGroupStartIdx - GroupSize] !=
+          StrideBetweenGroups)
+        return false;
+    }
 
-  // Iterate through all pointers and check if all distances are
-  // unique multiple of Dist.
-  SmallSet<int64_t, 4> Dists;
-  for (Value *Ptr : PointerOps) {
-    int64_t Dist = 0;
-    if (Ptr == PtrN)
-      Dist = Diff;
-    else if (Ptr != Ptr0)
-      Dist = *getPointersDiff(ScalarTy, Ptr0, ScalarTy, Ptr, *DL, *SE);
-    // If the strides are not the same or repeated, we can't
-    // vectorize.
-    if (((Dist / Stride) * Stride) != Dist || !Dists.insert(Dist).second)
-      break;
-  }
-  if (Dists.size() == Sz) {
-    Type *StrideTy = DL->getIndexType(Ptr0->getType());
-    SPtrInfo.StrideVal = ConstantInt::get(StrideTy, Stride);
-    SPtrInfo.Ty = getWidenedType(ScalarTy, Sz);
-    return true;
+    auto CheckGroup = [=](const unsigned StartIdx) -> bool {
+      auto Indices = seq<unsigned>(StartIdx + 1, Sz);
+      auto FoundIt = llvm::find_if(Indices, IsEndOfGroupIndex);
+      unsigned GroupEndIdx = FoundIt != Indices.end() ? *FoundIt : Sz;
+      return GroupEndIdx - StartIdx == GroupSize;
+    };
+    for (unsigned I = 0; I < Sz; I += GroupSize) {
+      if (!CheckGroup(I))
+        return false;
+    }
   }
-  return false;
+
+  Type *StrideTy = DL->getIndexType(Ptr0->getType());
+  SPtrInfo.StrideVal = ConstantInt::get(StrideTy, StrideIntVal);
+  SPtrInfo.Ty = getWidenedType(NewScalarTy, VecSz);
+  return true;
 }
 
 bool BoUpSLP::analyzeRtStrideCandidate(ArrayRef<Value *> PointerOps,
@@ -15061,11 +15121,19 @@ BoUpSLP::getEntryCost(const TreeEntry *E, ArrayRef<Value *> VectorizedVals,
         }
         break;
       case TreeEntry::StridedVectorize: {
+        const StridedPtrInfo &SPtrInfo = TreeEntryToStridedPtrInfoMap.at(E);
+        FixedVectorType *StridedLoadTy = SPtrInfo.Ty;
+        assert(StridedLoadTy && "Missing StridedPoinerInfo for tree entry.");
         Align CommonAlignment =
             computeCommonAlignment<LoadInst>(UniqueValues.getArrayRef());
         VecLdCost = TTI->getStridedMemoryOpCost(
-            Instruction::Load, VecTy, LI0->getPointerOperand(),
+            Instruction::Load, StridedLoadTy, LI0->getPointerOperand(),
             /*VariableMask=*/false, CommonAlignment, CostKind);
+        if (StridedLoadTy != VecTy)
+          VecLdCost +=
+              TTI->getCastInstrCost(Instruction::BitCast, VecTy, StridedLoadTy,
+                                    getCastContextHint(*E), CostKind);
+
         break;
       }
       case TreeEntry::CompressVectorize: {
@@ -19870,6 +19938,8 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
                      ? NewLI
                      : ::propagateMetadata(NewLI, E->Scalars);
 
+      if (StridedLoadTy != VecTy)
+        V = Builder.CreateBitOrPointerCast(V, VecTy);
       V = FinalShuffle(V, E);
       E->VectorizedValue = V;
       ++NumVectorInstructions;

llvm/test/Transforms/SLPVectorizer/RISCV/basic-strided-loads.ll

@@ -1,6 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
 
-; RUN: opt -mtriple=riscv64 -mattr=+m,+v -passes=slp-vectorizer -S < %s | FileCheck %s
+; RUN: opt -mtriple=riscv64 -mattr=+m,+v,+unaligned-vector-mem -passes=slp-vectorizer -S < %s | FileCheck %s
 
 define void @const_stride_1_no_reordering(ptr %pl, ptr %ps) {
 ; CHECK-LABEL: define void @const_stride_1_no_reordering(
@@ -621,21 +621,9 @@ define void @constant_stride_widen_no_reordering(ptr %pl, i64 %stride, ptr %ps)
 ; CHECK-LABEL: define void @constant_stride_widen_no_reordering(
 ; CHECK-SAME: ptr [[PL:%.*]], i64 [[STRIDE:%.*]], ptr [[PS:%.*]]) #[[ATTR0]] {
 ; CHECK-NEXT:    [[GEP_L0:%.*]] = getelementptr inbounds i8, ptr [[PL]], i64 0
-; CHECK-NEXT:    [[GEP_L4:%.*]] = getelementptr inbounds i8, ptr [[PL]], i64 100
-; CHECK-NEXT:    [[GEP_L8:%.*]] = getelementptr inbounds i8, ptr [[PL]], i64 200
-; CHECK-NEXT:    [[GEP_L12:%.*]] = getelementptr inbounds i8, ptr [[PL]], i64 300
 ; CHECK-NEXT:    [[GEP_S0:%.*]] = getelementptr inbounds i8, ptr [[PS]], i64 0
-; CHECK-NEXT:    [[TMP1:%.*]] = load <4 x i8>, ptr [[GEP_L0]], align 1
-; CHECK-NEXT:    [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_L4]], align 1
-; CHECK-NEXT:    [[TMP3:%.*]] = load <4 x i8>, ptr [[GEP_L8]], align 1
-; CHECK-NEXT:    [[TMP4:%.*]] = load <4 x i8>, ptr [[GEP_L12]], align 1
-; CHECK-NEXT:    [[TMP5:%.*]] = shufflevector <4 x i8> [[TMP1]], <4 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
-; CHECK-NEXT:    [[TMP6:%.*]] = shufflevector <4 x i8> [[TMP2]], <4 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
-; CHECK-NEXT:    [[TMP7:%.*]] = shufflevector <4 x i8> [[TMP1]], <4 x i8> [[TMP2]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
-; CHECK-NEXT:    [[TMP11:%.*]] = shufflevector <4 x i8> [[TMP3]], <4 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
-; CHECK-NEXT:    [[TMP9:%.*]] = shufflevector <16 x i8> [[TMP7]], <16 x i8> [[TMP11]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 16, i32 17, i32 18, i32 19, i32 poison, i32 poison, i32 poison, i32 poison>
-; CHECK-NEXT:    [[TMP10:%.*]] = shufflevector <4 x i8> [[TMP4]], <4 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
-; CHECK-NEXT:    [[TMP8:%.*]] = shufflevector <16 x i8> [[TMP9]], <16 x i8> [[TMP10]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 16, i32 17, i32 18, i32 19>
+; CHECK-NEXT:    [[TMP1:%.*]] = call <4 x i32> @llvm.experimental.vp.strided.load.v4i32.p0.i64(ptr align 1 [[GEP_L0]], i64 100, <4 x i1> splat (i1 true), i32 4)
+; CHECK-NEXT:    [[TMP8:%.*]] = bitcast <4 x i32> [[TMP1]] to <16 x i8>
 ; CHECK-NEXT:    store <16 x i8> [[TMP8]], ptr [[GEP_S0]], align 1
 ; CHECK-NEXT:    ret void
 ;