The third conversion method for vp.reverse

It may temporarily lose some performance when EVL tail folding.

Author: Mel-Chen

llvm/lib/Transforms/Vectorize/VPlan.h

@@ -3484,6 +3484,15 @@ struct VPWidenStoreEVLRecipe final : public VPWidenMemoryRecipe {
     setMask(Mask);
   }
 
+  VPWidenStoreEVLRecipe(VPWidenStoreRecipe &S, VPValue *Addr,
+                        VPValue *StoredVal, VPValue &EVL, VPValue *Mask)
+      : VPWidenMemoryRecipe(VPDef::VPWidenStoreEVLSC, S.getIngredient(),
+                            {Addr, StoredVal, &EVL}, S.isConsecutive(),
+                            S.isReverse(), S, S.getDebugLoc()) {
+    assert(isReverse() && "Only reverse access need to set new stored value");
+    setMask(Mask);
+  }
+
   VP_CLASSOF_IMPL(VPDef::VPWidenStoreEVLSC)
 
   /// Return the address accessed by this recipe.

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

@@ -2856,25 +2856,43 @@ static VPRecipeBase *optimizeMaskToEVL(VPValue *HeaderMask,
     return new VPWidenLoadEVLRecipe(cast<VPWidenLoadRecipe>(CurRecipe), Addr,
                                     EVL, Mask);
 
-  if (match(&CurRecipe,
+  VPValue *ReversedVal;
+  if (match(&CurRecipe, m_Reverse(m_VPValue(ReversedVal))) &&
+      match(ReversedVal,
             m_MaskedLoad(m_VPValue(EndPtr), m_RemoveMask(HeaderMask, Mask))) &&
       match(EndPtr, m_VecEndPtr(m_VPValue(Addr), m_Specific(&Plan->getVF()))) &&
-      cast<VPWidenLoadRecipe>(CurRecipe).isReverse())
-    return new VPWidenLoadEVLRecipe(cast<VPWidenLoadRecipe>(CurRecipe),
-                                    AdjustEndPtr(EndPtr), EVL, Mask);
+      cast<VPWidenLoadRecipe>(ReversedVal)->isReverse()) {
+    auto *LoadR = new VPWidenLoadEVLRecipe(
+        *cast<VPWidenLoadRecipe>(ReversedVal), AdjustEndPtr(EndPtr), EVL, Mask);
+    LoadR->insertBefore(&CurRecipe);
+    return new VPWidenIntrinsicRecipe(
+        Intrinsic::experimental_vp_reverse, {LoadR, Plan->getTrue(), &EVL},
+        TypeInfo.inferScalarType(LoadR), {}, {}, DL);
+  }
 
   if (match(&CurRecipe, m_MaskedStore(m_VPValue(Addr), m_VPValue(),
                                       m_RemoveMask(HeaderMask, Mask))) &&
       !cast<VPWidenStoreRecipe>(CurRecipe).isReverse())
     return new VPWidenStoreEVLRecipe(cast<VPWidenStoreRecipe>(CurRecipe), Addr,
                                      EVL, Mask);
 
-  if (match(&CurRecipe, m_MaskedStore(m_VPValue(EndPtr), m_VPValue(),
+  VPValue *StoredVal;
+  if (match(&CurRecipe, m_MaskedStore(m_VPValue(EndPtr), m_VPValue(StoredVal),
                                       m_RemoveMask(HeaderMask, Mask))) &&
       match(EndPtr, m_VecEndPtr(m_VPValue(Addr), m_Specific(&Plan->getVF()))) &&
-      cast<VPWidenStoreRecipe>(CurRecipe).isReverse())
-    return new VPWidenStoreEVLRecipe(cast<VPWidenStoreRecipe>(CurRecipe),
-                                     AdjustEndPtr(EndPtr), EVL, Mask);
+      cast<VPWidenStoreRecipe>(CurRecipe).isReverse()) {
+    if (match(StoredVal, m_Reverse(m_VPValue(ReversedVal)))) {
+      auto *NewReverse = new VPWidenIntrinsicRecipe(
+          Intrinsic::experimental_vp_reverse,
+          {ReversedVal, Plan->getTrue(), &EVL},
+          TypeInfo.inferScalarType(ReversedVal), {}, {},
+          cast<VPInstruction>(StoredVal)->getDebugLoc());
+      NewReverse->insertBefore(&CurRecipe);
+      return new VPWidenStoreEVLRecipe(cast<VPWidenStoreRecipe>(CurRecipe),
+                                       AdjustEndPtr(EndPtr), NewReverse, EVL,
+                                       Mask);
+    }
+  }
 
   if (auto *Rdx = dyn_cast<VPReductionRecipe>(&CurRecipe))
     if (Rdx->isConditional() &&
@@ -2947,7 +2965,6 @@ static void transformRecipestoEVLRecipes(VPlan &Plan, VPValue &EVL) {
   // contained.
   bool ContainsFORs =
       any_of(Header->phis(), IsaPred<VPFirstOrderRecurrencePHIRecipe>);
-  VPValue *PrevEVL = nullptr;
   if (ContainsFORs) {
     // TODO: Use VPInstruction::ExplicitVectorLength to get maximum EVL.
     VPValue *MaxEVL = &Plan.getVF();
@@ -2958,42 +2975,28 @@ static void transformRecipestoEVLRecipes(VPlan &Plan, VPValue &EVL) {
         TypeInfo.inferScalarType(MaxEVL), DebugLoc::getUnknown());
 
     Builder.setInsertPoint(Header, Header->getFirstNonPhi());
-    PrevEVL = Builder.createScalarPhi({MaxEVL, &EVL}, DebugLoc::getUnknown(),
-                                      "prev.evl");
-  }
-
-  // Transform the recipes must be converted to vector predication intrinsics
-  // even if they do not use header mask.
-  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
-           vp_depth_first_deep(Plan.getVectorLoopRegion()->getEntry()))) {
-    for (VPRecipeBase &R : *VPBB) {
-      VPWidenIntrinsicRecipe *NewRecipe = nullptr;
-      VPValue *V1, *V2;
-      if (match(&R, m_VPInstruction<VPInstruction::FirstOrderRecurrenceSplice>(
-                        m_VPValue(V1), m_VPValue(V2)))) {
+    VPValue *PrevEVL = Builder.createScalarPhi(
+        {MaxEVL, &EVL}, DebugLoc::getUnknown(), "prev.evl");
+
+    for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
+             vp_depth_first_deep(Plan.getVectorLoopRegion()->getEntry()))) {
+      for (VPRecipeBase &R : *VPBB) {
+        VPValue *V1, *V2;
+        if (!match(&R,
+                   m_VPInstruction<VPInstruction::FirstOrderRecurrenceSplice>(
+                       m_VPValue(V1), m_VPValue(V2))))
+          continue;
         VPValue *Imm = Plan.getOrAddLiveIn(
             ConstantInt::getSigned(Type::getInt32Ty(Plan.getContext()), -1));
-        NewRecipe = new VPWidenIntrinsicRecipe(
+        VPWidenIntrinsicRecipe *VPSplice = new VPWidenIntrinsicRecipe(
             Intrinsic::experimental_vp_splice,
             {V1, V2, Imm, Plan.getTrue(), PrevEVL, &EVL},
             TypeInfo.inferScalarType(R.getVPSingleValue()), {}, {},
             R.getDebugLoc());
+        VPSplice->insertBefore(&R);
+        R.getVPSingleValue()->replaceAllUsesWith(VPSplice);
+        ToErase.push_back(&R);
       }
-
-      // TODO: Only convert reverse to vp.reverse if it uses the result of
-      // vp.load, or defines the stored value of vp.store.
-      if (match(&R, m_Reverse(m_VPValue(V1)))) {
-        NewRecipe = new VPWidenIntrinsicRecipe(
-            Intrinsic::experimental_vp_reverse, {V1, Plan.getTrue(), &EVL},
-            TypeInfo.inferScalarType(R.getVPSingleValue()), {}, {},
-            R.getDebugLoc());
-      }
-
-      if (!NewRecipe)
-        continue;
-      NewRecipe->insertBefore(&R);
-      R.getVPSingleValue()->replaceAllUsesWith(NewRecipe);
-      ToErase.push_back(&R);
     }
   }
 

llvm/test/Transforms/LoopVectorize/RISCV/tail-folding-uniform-store.ll

@@ -15,7 +15,6 @@ define void @lshift_significand(i32 %n, ptr nocapture writeonly %dst) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = sub i64 3, [[SPEC_SELECT]]
 ; CHECK-NEXT:    br label %[[VECTOR_PH:.*]]
 ; CHECK:       [[VECTOR_PH]]:
-; CHECK-NEXT:    [[REVERSE:%.*]] = call <vscale x 2 x i64> @llvm.vector.reverse.nxv2i64(<vscale x 2 x i64> zeroinitializer)
 ; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
 ; CHECK:       [[VECTOR_BODY]]:
 ; CHECK-NEXT:    [[EVL_BASED_IV:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_EVL_NEXT:%.*]], %[[VECTOR_BODY]] ]
@@ -24,6 +23,7 @@ define void @lshift_significand(i32 %n, ptr nocapture writeonly %dst) {
 ; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = add i64 [[SPEC_SELECT]], [[EVL_BASED_IV]]
 ; CHECK-NEXT:    [[TMP12:%.*]] = sub nuw nsw i64 1, [[OFFSET_IDX]]
 ; CHECK-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr i64, ptr [[DST]], i64 [[TMP12]]
+; CHECK-NEXT:    [[REVERSE:%.*]] = call <vscale x 2 x i64> @llvm.experimental.vp.reverse.nxv2i64(<vscale x 2 x i64> zeroinitializer, <vscale x 2 x i1> splat (i1 true), i32 [[TMP1]])
 ; CHECK-NEXT:    [[TMP4:%.*]] = zext i32 [[TMP1]] to i64
 ; CHECK-NEXT:    [[TMP5:%.*]] = mul i64 0, [[TMP4]]
 ; CHECK-NEXT:    [[TMP6:%.*]] = sub i64 [[TMP4]], 1