buildTree_rec(): Single point of definition for legality characteristics

Author: gbossu

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -4248,13 +4248,34 @@ class BoUpSLP {
   bool areAltOperandsProfitable(const InstructionsState &S,
                                 ArrayRef<Value *> VL) const;
 
+  /// Contains all the outputs of legality analysis for a list of values to
+  /// vectorize.
+  class ScalarsVectorizationLegality {
+    InstructionsState S;
+    bool IsLegal;
+    bool TryToFindDuplicates;
+    bool TrySplitVectorize;
+
+  public:
+    ScalarsVectorizationLegality(InstructionsState S, bool IsLegal,
+                                 bool TryToFindDuplicates = true,
+                                 bool TrySplitVectorize = false)
+        : S(S), IsLegal(IsLegal), TryToFindDuplicates(TryToFindDuplicates),
+          TrySplitVectorize(TrySplitVectorize) {
+      assert((!IsLegal || (S.valid() && TryToFindDuplicates)) &&
+             "Inconsistent state");
+    }
+    const InstructionsState &getInstructionsState() const { return S; };
+    bool isLegal() const { return IsLegal; }
+    bool tryToFindDuplicates() const { return TryToFindDuplicates; }
+    bool trySplitVectorize() const { return TrySplitVectorize; }
+  };
+
   /// Checks if the specified list of the instructions/values can be vectorized
   /// in general.
-  bool isLegalToVectorizeScalars(ArrayRef<Value *> VL, unsigned Depth,
-                                 const EdgeInfo &UserTreeIdx,
-                                 InstructionsState &S,
-                                 bool &TryToFindDuplicates,
-                                 bool &TrySplitVectorize) const;
+  ScalarsVectorizationLegality
+  getScalarsVectorizationLegality(ArrayRef<Value *> VL, unsigned Depth,
+                                  const EdgeInfo &UserTreeIdx) const;
 
   /// Checks if the specified list of the instructions/values can be vectorized
   /// and fills required data before actual scheduling of the instructions.
@@ -9693,25 +9714,21 @@ bool BoUpSLP::canBuildSplitNode(ArrayRef<Value *> VL,
   return true;
 }
 
-bool BoUpSLP::isLegalToVectorizeScalars(ArrayRef<Value *> VL, unsigned Depth,
-                                        const EdgeInfo &UserTreeIdx,
-                                        InstructionsState &S,
-                                        bool &TryToFindDuplicates,
-                                        bool &TrySplitVectorize) const {
+BoUpSLP::ScalarsVectorizationLegality
+BoUpSLP::getScalarsVectorizationLegality(ArrayRef<Value *> VL, unsigned Depth,
+                                         const EdgeInfo &UserTreeIdx) const {
   assert((allConstant(VL) || allSameType(VL)) && "Invalid types!");
 
-  S = getSameOpcode(VL, *TLI);
-  TryToFindDuplicates = true;
-  TrySplitVectorize = false;
+  InstructionsState S = getSameOpcode(VL, *TLI);
 
   // Don't go into catchswitch blocks, which can happen with PHIs.
   // Such blocks can only have PHIs and the catchswitch.  There is no
   // place to insert a shuffle if we need to, so just avoid that issue.
   if (S && isa<CatchSwitchInst>(S.getMainOp()->getParent()->getTerminator())) {
     LLVM_DEBUG(dbgs() << "SLP: bundle in catchswitch block.\n");
     // Do not try to pack to avoid extra instructions here.
-    TryToFindDuplicates = false;
-    return false;
+    return ScalarsVectorizationLegality(S, /*IsLegal=*/false,
+                                        /*TryToFindDuplicates=*/false);
   }
 
   // Check if this is a duplicate of another entry.
@@ -9721,14 +9738,14 @@ bool BoUpSLP::isLegalToVectorizeScalars(ArrayRef<Value *> VL, unsigned Depth,
       if (E->isSame(VL)) {
         LLVM_DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *S.getMainOp()
                           << ".\n");
-        return false;
+        return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
       }
       SmallPtrSet<Value *, 8> Values(llvm::from_range, E->Scalars);
       if (all_of(VL, [&](Value *V) {
             return isa<PoisonValue>(V) || Values.contains(V);
           })) {
         LLVM_DEBUG(dbgs() << "SLP: Gathering due to full overlap.\n");
-        return false;
+        return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
       }
     }
   }
@@ -9745,23 +9762,23 @@ bool BoUpSLP::isLegalToVectorizeScalars(ArrayRef<Value *> VL, unsigned Depth,
                   cast<Instruction>(I)->getOpcode() == S.getOpcode();
          })))) {
     LLVM_DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n");
-    return false;
+    return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
   }
 
   // Don't handle scalable vectors
   if (S && S.getOpcode() == Instruction::ExtractElement &&
       isa<ScalableVectorType>(
           cast<ExtractElementInst>(S.getMainOp())->getVectorOperandType())) {
     LLVM_DEBUG(dbgs() << "SLP: Gathering due to scalable vector type.\n");
-    return false;
+    return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
   }
 
   // Don't handle vectors.
   if (!SLPReVec && getValueType(VL.front())->isVectorTy()) {
     LLVM_DEBUG(dbgs() << "SLP: Gathering due to vector type.\n");
     // Do not try to pack to avoid extra instructions here.
-    TryToFindDuplicates = false;
-    return false;
+    return ScalarsVectorizationLegality(S, /*IsLegal=*/false,
+                                        /*TryToFindDuplicates=*/false);
   }
 
   // If all of the operands are identical or constant we have a simple solution.
@@ -9851,11 +9868,12 @@ bool BoUpSLP::isLegalToVectorizeScalars(ArrayRef<Value *> VL, unsigned Depth,
     if (!S) {
       LLVM_DEBUG(dbgs() << "SLP: Try split and if failed, gathering due to "
                            "C,S,B,O, small shuffle. \n");
-      TrySplitVectorize = true;
-      return false;
+      return ScalarsVectorizationLegality(S, /*IsLegal=*/false,
+                                          /*TryToFindDuplicates=*/true,
+                                          /*TrySplitVectorize=*/true);
     }
     LLVM_DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O, small shuffle. \n");
-    return false;
+    return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
   }
 
   // Don't vectorize ephemeral values.
@@ -9865,8 +9883,8 @@ bool BoUpSLP::isLegalToVectorizeScalars(ArrayRef<Value *> VL, unsigned Depth,
         LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *V
                           << ") is ephemeral.\n");
         // Do not try to pack to avoid extra instructions here.
-        TryToFindDuplicates = false;
-        return false;
+        return ScalarsVectorizationLegality(S, /*IsLegal=*/false,
+                                            /*TryToFindDuplicates=*/false);
       }
     }
   }
@@ -9915,7 +9933,7 @@ bool BoUpSLP::isLegalToVectorizeScalars(ArrayRef<Value *> VL, unsigned Depth,
     if (PreferScalarize) {
       LLVM_DEBUG(dbgs() << "SLP: The instructions are in tree and alternate "
                            "node is not profitable.\n");
-      return false;
+      return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
     }
   }
 
@@ -9924,7 +9942,7 @@ bool BoUpSLP::isLegalToVectorizeScalars(ArrayRef<Value *> VL, unsigned Depth,
     for (Value *V : VL) {
       if (UserIgnoreList->contains(V)) {
         LLVM_DEBUG(dbgs() << "SLP: Gathering due to gathered scalar.\n");
-        return false;
+        return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
       }
     }
   }
@@ -9954,9 +9972,9 @@ bool BoUpSLP::isLegalToVectorizeScalars(ArrayRef<Value *> VL, unsigned Depth,
     // Do not vectorize EH and non-returning blocks, not profitable in most
     // cases.
     LLVM_DEBUG(dbgs() << "SLP: bundle in unreachable block.\n");
-    return false;
+    return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
   }
-  return true;
+  return ScalarsVectorizationLegality(S, /*IsLegal=*/true);
 }
 
 void BoUpSLP::buildTree_rec(ArrayRef<Value *> VLRef, unsigned Depth,
@@ -9973,7 +9991,6 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VLRef, unsigned Depth,
                  ReuseShuffleIndices);
   };
 
-  InstructionsState S = InstructionsState::invalid();
   // Tries to build split node.
   auto TrySplitNode = [&](const InstructionsState &LocalState) {
     SmallVector<Value *> Op1, Op2;
@@ -10007,17 +10024,17 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VLRef, unsigned Depth,
     return true;
   };
 
-  bool TryToPackDuplicates;
-  bool TrySplitVectorize;
-  if (!isLegalToVectorizeScalars(VL, Depth, UserTreeIdx, S, TryToPackDuplicates,
-                                 TrySplitVectorize)) {
-    if (TrySplitVectorize) {
+  ScalarsVectorizationLegality SVL =
+      getScalarsVectorizationLegality(VL, Depth, UserTreeIdx);
+  const InstructionsState &S = SVL.getInstructionsState();
+  if (!SVL.isLegal()) {
+    if (SVL.trySplitVectorize()) {
       auto [MainOp, AltOp] = getMainAltOpsNoStateVL(VL);
       // Last chance to try to vectorize alternate node.
       if (MainOp && AltOp && TrySplitNode(InstructionsState(MainOp, AltOp)))
         return;
     }
-    if (TryToPackDuplicates)
+    if (SVL.tryToFindDuplicates())
       tryToFindDuplicates(VL, ReuseShuffleIndices, *TTI, *TLI, S, UserTreeIdx);
 
     CreateGatherTreeEntry(S);