[LV][EVL]Support reversed loads/stores.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -1571,13 +1571,7 @@ class LoopVectorizationCostModel {
   /// Returns true if VP intrinsics with explicit vector length support should
   /// be generated in the tail folded loop.
   bool foldTailWithEVL() const {
-    return getTailFoldingStyle() == TailFoldingStyle::DataWithEVL &&
-           // FIXME: remove this once vp_reverse is supported.
-           none_of(
-               WideningDecisions,
-               [](const std::pair<std::pair<Instruction *, ElementCount>,
-                                  std::pair<InstWidening, InstructionCost>>
-                      &Data) { return Data.second.first == CM_Widen_Reverse; });
+    return getTailFoldingStyle() == TailFoldingStyle::DataWithEVL;
   }
 
   /// Returns true if the Phi is part of an inloop reduction.
@@ -9383,12 +9377,18 @@ void VPWidenLoadRecipe::execute(VPTransformState &State) {
   }
 }
 
+static Instruction *createReverseEVL(IRBuilderBase &Builder, Value *Operand,
+                                     Value *EVL, const Twine &Name) {
+  VectorType *ValTy = cast<VectorType>(Operand->getType());
+  Value *AllTrueMask =
+      Builder.CreateVectorSplat(ValTy->getElementCount(), Builder.getTrue());
+  return Builder.CreateIntrinsic(ValTy, Intrinsic::experimental_vp_reverse,
+                                 {Operand, AllTrueMask, EVL}, nullptr, Name);
+}
+
 void VPWidenLoadEVLRecipe::execute(VPTransformState &State) {
   assert(State.UF == 1 && "Expected only UF == 1 when vectorizing with "
                           "explicit vector length.");
-  // FIXME: Support reverse loading after vp_reverse is added.
-  assert(!isReverse() && "Reverse loads are not implemented yet.");
-
   auto *LI = cast<LoadInst>(&Ingredient);
 
   Type *ScalarDataTy = getLoadStoreType(&Ingredient);
@@ -9401,9 +9401,15 @@ void VPWidenLoadEVLRecipe::execute(VPTransformState &State) {
   CallInst *NewLI;
   Value *EVL = State.get(getEVL(), VPIteration(0, 0));
   Value *Addr = State.get(getAddr(), 0, !CreateGather);
-  Value *Mask = getMask()
-                    ? State.get(getMask(), 0)
-                    : Builder.CreateVectorSplat(State.VF, Builder.getTrue());
+  Value *Mask = nullptr;
+  if (VPValue *VPMask = getMask()) {
+    Mask = State.get(VPMask, 0);
+    if (isReverse())
+      Mask = createReverseEVL(Builder, Mask, EVL, "vp.reverse.mask");
+  } else {
+    Mask = Builder.CreateVectorSplat(State.VF, Builder.getTrue());
+  }
+
   if (CreateGather) {
     NewLI =
         Builder.CreateIntrinsic(DataTy, Intrinsic::vp_gather, {Addr, Mask, EVL},
@@ -9417,7 +9423,13 @@ void VPWidenLoadEVLRecipe::execute(VPTransformState &State) {
   NewLI->addParamAttr(
       0, Attribute::getWithAlignment(NewLI->getContext(), Alignment));
   State.addMetadata(NewLI, LI);
-  State.set(this, NewLI, 0);
+  Instruction *Res = NewLI;
+  if (isReverse()) {
+    // Use cheap all-true mask for reverse rather than actual mask, it does not
+    // affect the result.
+    Res = createReverseEVL(Builder, Res, EVL, "vp.reverse");
+  }
+  State.set(this, Res, 0);
 }
 
 void VPWidenStoreRecipe::execute(VPTransformState &State) {
@@ -9463,9 +9475,6 @@ void VPWidenStoreRecipe::execute(VPTransformState &State) {
 void VPWidenStoreEVLRecipe::execute(VPTransformState &State) {
   assert(State.UF == 1 && "Expected only UF == 1 when vectorizing with "
                           "explicit vector length.");
-  // FIXME: Support reverse loading after vp_reverse is added.
-  assert(!isReverse() && "Reverse store are not implemented yet.");
-
   auto *SI = cast<StoreInst>(&Ingredient);
 
   VPValue *StoredValue = getStoredValue();
@@ -9478,10 +9487,19 @@ void VPWidenStoreEVLRecipe::execute(VPTransformState &State) {
   CallInst *NewSI = nullptr;
   Value *StoredVal = State.get(StoredValue, 0);
   Value *EVL = State.get(getEVL(), VPIteration(0, 0));
-  // FIXME: Support reverse store after vp_reverse is added.
-  Value *Mask = getMask()
-                    ? State.get(getMask(), 0)
-                    : Builder.CreateVectorSplat(State.VF, Builder.getTrue());
+  if (isReverse()) {
+    // Use cheap all-true mask for reverse rather than actual mask, it does not
+    // affect the result.
+    StoredVal = createReverseEVL(Builder, StoredVal, EVL, "vp.reverse");
+  }
+  Value *Mask = nullptr;
+  if (VPValue *VPMask = getMask()) {
+    Mask = State.get(VPMask, 0);
+    if (isReverse())
+      Mask = createReverseEVL(Builder, Mask, EVL, "vp.reverse.mask");
+  } else {
+    Mask = Builder.CreateVectorSplat(State.VF, Builder.getTrue());
+  }
   Value *Addr = State.get(getAddr(), 0, !CreateScatter);
   if (CreateScatter) {
     NewSI = Builder.CreateIntrinsic(Type::getVoidTy(EVL->getContext()),

llvm/lib/Transforms/Vectorize/VPlan.h

@@ -2399,8 +2399,8 @@ struct VPWidenLoadRecipe final : public VPWidenMemoryRecipe, public VPValue {
 struct VPWidenLoadEVLRecipe final : public VPWidenMemoryRecipe, public VPValue {
   VPWidenLoadEVLRecipe(VPWidenLoadRecipe *L, VPValue *EVL, VPValue *Mask)
       : VPWidenMemoryRecipe(VPDef::VPWidenLoadEVLSC, L->getIngredient(),
-                            {L->getAddr(), EVL}, L->isConsecutive(), false,
-                            L->getDebugLoc()),
+                            {L->getAddr(), EVL}, L->isConsecutive(),
+                            L->isReverse(), L->getDebugLoc()),
         VPValue(this, &getIngredient()) {
     setMask(Mask);
   }
@@ -2476,7 +2476,8 @@ struct VPWidenStoreEVLRecipe final : public VPWidenMemoryRecipe {
   VPWidenStoreEVLRecipe(VPWidenStoreRecipe *S, VPValue *EVL, VPValue *Mask)
       : VPWidenMemoryRecipe(VPDef::VPWidenStoreEVLSC, S->getIngredient(),
                             {S->getAddr(), S->getStoredValue(), EVL},
-                            S->isConsecutive(), false, S->getDebugLoc()) {
+                            S->isConsecutive(), S->isReverse(),
+                            S->getDebugLoc()) {
     setMask(Mask);
   }
 

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

@@ -1341,8 +1341,6 @@ void VPlanTransforms::addExplicitVectorLength(VPlan &Plan) {
       auto *MemR = dyn_cast<VPWidenMemoryRecipe>(U);
       if (!MemR)
         continue;
-      assert(!MemR->isReverse() &&
-             "Reversed memory operations not supported yet.");
       VPValue *OrigMask = MemR->getMask();
       assert(OrigMask && "Unmasked widen memory recipe when folding tail");
       VPValue *NewMask = HeaderMask == OrigMask ? nullptr : OrigMask;