[SLP]Improve minbitwidth analysis for operands of IToFP and ICmp instructions.

Compiler can improve analysis for operands of UIToFP/SIToFP instructions
and operands of ICmp instruction.

Reviewers: RKSimon

Reviewed By: RKSimon

Pull Request: #85966

Author: alexey-bataev

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -1107,7 +1107,7 @@ class BoUpSLP {
     MinBWs.clear();
     ReductionBitWidth = 0;
     CastMaxMinBWSizes.reset();
-    TruncNodes.clear();
+    ExtraBitWidthNodes.clear();
     InstrElementSize.clear();
     UserIgnoreList = nullptr;
     PostponedGathers.clear();
@@ -3683,8 +3683,9 @@ class BoUpSLP {
   /// type sizes, used in the tree.
   std::optional<std::pair<unsigned, unsigned>> CastMaxMinBWSizes;
 
-  /// Indices of the vectorized trunc nodes.
-  DenseSet<unsigned> TruncNodes;
+  /// Indices of the vectorized nodes, which supposed to be the roots of the new
+  /// bitwidth analysis attempt, like trunc, IToFP or ICmp.
+  DenseSet<unsigned> ExtraBitWidthNodes;
 };
 
 } // end namespace slpvectorizer
@@ -6612,7 +6613,18 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
                 PrevMaxBW),
             std::min<unsigned>(DL->getTypeSizeInBits(VL0->getType()),
                                PrevMinBW));
-        TruncNodes.insert(VectorizableTree.size());
+        ExtraBitWidthNodes.insert(VectorizableTree.size() + 1);
+      } else if (ShuffleOrOp == Instruction::SIToFP ||
+                 ShuffleOrOp == Instruction::UIToFP) {
+        unsigned NumSignBits =
+             ComputeNumSignBits(VL0->getOperand(0), *DL, 0, AC, nullptr, DT);
+        if (auto *OpI = dyn_cast<Instruction>(VL0->getOperand(0))) {
+          APInt Mask = DB->getDemandedBits(OpI);
+          NumSignBits = std::max(NumSignBits, Mask.countl_zero());
+        }
+        if (NumSignBits * 2 >=
+            DL->getTypeSizeInBits(VL0->getOperand(0)->getType()))
+          ExtraBitWidthNodes.insert(VectorizableTree.size() + 1);
       }
       TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                    ReuseShuffleIndicies);
@@ -6660,6 +6672,18 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
       TE->setOperand(1, Right);
       buildTree_rec(Left, Depth + 1, {TE, 0});
       buildTree_rec(Right, Depth + 1, {TE, 1});
+      if (ShuffleOrOp == Instruction::ICmp) {
+        unsigned NumSignBits0 =
+            ComputeNumSignBits(VL0->getOperand(0), *DL, 0, AC, nullptr, DT);
+        if (NumSignBits0 * 2 >=
+            DL->getTypeSizeInBits(VL0->getOperand(0)->getType()))
+          ExtraBitWidthNodes.insert(getOperandEntry(TE, 0)->Idx);
+        unsigned NumSignBits1 =
+            ComputeNumSignBits(VL0->getOperand(1), *DL, 0, AC, nullptr, DT);
+        if (NumSignBits1 * 2 >=
+            DL->getTypeSizeInBits(VL0->getOperand(1)->getType()))
+          ExtraBitWidthNodes.insert(getOperandEntry(TE, 1)->Idx);
+      }
       return;
     }
     case Instruction::Select:
@@ -14302,7 +14326,8 @@ void BoUpSLP::computeMinimumValueSizes() {
   bool IsStoreOrInsertElt =
       VectorizableTree.front()->getOpcode() == Instruction::Store ||
       VectorizableTree.front()->getOpcode() == Instruction::InsertElement;
-  if ((IsStoreOrInsertElt || UserIgnoreList) && TruncNodes.size() <= 1 &&
+  if ((IsStoreOrInsertElt || UserIgnoreList) &&
+      ExtraBitWidthNodes.size() <= 1 &&
       (!CastMaxMinBWSizes || CastMaxMinBWSizes->second == 0 ||
        CastMaxMinBWSizes->first / CastMaxMinBWSizes->second <= 2))
     return;
@@ -14506,16 +14531,21 @@ void BoUpSLP::computeMinimumValueSizes() {
     IsTopRoot = false;
     IsProfitableToDemoteRoot = true;
 
-    if (TruncNodes.empty()) {
+    if (ExtraBitWidthNodes.empty()) {
       NodeIdx = VectorizableTree.size();
     } else {
       unsigned NewIdx = 0;
       do {
-        NewIdx = *TruncNodes.begin() + 1;
-        TruncNodes.erase(TruncNodes.begin());
-      } while (NewIdx <= NodeIdx && !TruncNodes.empty());
+        NewIdx = *ExtraBitWidthNodes.begin();
+        ExtraBitWidthNodes.erase(ExtraBitWidthNodes.begin());
+      } while (NewIdx <= NodeIdx && !ExtraBitWidthNodes.empty());
       NodeIdx = NewIdx;
-      IsTruncRoot = true;
+      IsTruncRoot = any_of(
+          VectorizableTree[NewIdx]->UserTreeIndices, [](const EdgeInfo &EI) {
+            return EI.EdgeIdx == 0 &&
+                   EI.UserTE->getOpcode() == Instruction::ICmp &&
+                   !EI.UserTE->isAltShuffle();
+          });
     }
 
     // If the maximum bit width we compute is less than the with of the roots'

llvm/test/Transforms/SLPVectorizer/X86/minbitwidth-icmp-to-trunc.ll

@@ -19,8 +19,8 @@ define i1 @test(ptr noalias %0, i64 %1, ptr noalias %p, ptr %p1) {
 ; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq <2 x i24> [[TMP8]], <i24 24, i24 24>
 ; CHECK-NEXT:    [[TMP10:%.*]] = select <2 x i1> [[TMP9]], <2 x i24> <i24 23, i24 23>, <2 x i24> [[TMP8]]
 ; CHECK-NEXT:    [[TMP23:%.*]] = trunc <2 x i24> [[TMP10]] to <2 x i8>
-; CHECK-NEXT:    [[TMP11:%.*]] = zext <2 x i8> [[TMP23]] to <2 x i32>
-; CHECK-NEXT:    [[TMP12:%.*]] = and <2 x i32> [[TMP11]], <i32 254, i32 254>
+; CHECK-NEXT:    [[TMP26:%.*]] = zext <2 x i8> [[TMP23]] to <2 x i32>
+; CHECK-NEXT:    [[TMP12:%.*]] = and <2 x i32> [[TMP26]], <i32 254, i32 254>
 ; CHECK-NEXT:    [[TMP13:%.*]] = icmp eq <2 x i32> [[TMP12]], <i32 4, i32 4>
 ; CHECK-NEXT:    [[TMP25:%.*]] = select <2 x i1> [[TMP13]], <2 x i8> <i8 2, i8 2>, <2 x i8> [[TMP23]]
 ; CHECK-NEXT:    [[TMP14:%.*]] = zext <2 x i8> [[TMP25]] to <2 x i32>

llvm/test/Transforms/SLPVectorizer/X86/minbitwidth-node-with-multi-users.ll

@@ -10,12 +10,14 @@ define void @test() {
 ; CHECK-NEXT:    [[TMP3:%.*]] = select i1 false, i32 0, i32 0
 ; CHECK-NEXT:    [[TMP4:%.*]] = insertelement <4 x i8> <i8 poison, i8 0, i8 poison, i8 poison>, i8 [[TMP1]], i32 0
 ; CHECK-NEXT:    [[TMP5:%.*]] = shufflevector <4 x i8> [[TMP4]], <4 x i8> poison, <4 x i32> <i32 0, i32 0, i32 0, i32 1>
-; CHECK-NEXT:    [[TMP6:%.*]] = sext <4 x i8> [[TMP5]] to <4 x i32>
+; CHECK-NEXT:    [[TMP15:%.*]] = trunc <4 x i8> [[TMP5]] to <4 x i1>
 ; CHECK-NEXT:    [[TMP7:%.*]] = shufflevector <4 x i8> [[TMP4]], <4 x i8> poison, <4 x i32> zeroinitializer
 ; CHECK-NEXT:    [[TMP8:%.*]] = or <4 x i8> [[TMP7]], zeroinitializer
-; CHECK-NEXT:    [[TMP9:%.*]] = sext <4 x i8> [[TMP8]] to <4 x i32>
-; CHECK-NEXT:    [[TMP10:%.*]] = or <4 x i32> zeroinitializer, [[TMP6]]
-; CHECK-NEXT:    [[TMP11:%.*]] = icmp eq <4 x i32> [[TMP9]], [[TMP10]]
+; CHECK-NEXT:    [[TMP9:%.*]] = trunc <4 x i8> [[TMP8]] to <4 x i1>
+; CHECK-NEXT:    [[TMP10:%.*]] = or <4 x i1> zeroinitializer, [[TMP15]]
+; CHECK-NEXT:    [[TMP11:%.*]] = icmp eq <4 x i1> [[TMP9]], [[TMP10]]
+; CHECK-NEXT:    [[TMP16:%.*]] = shufflevector <4 x i1> [[TMP15]], <4 x i1> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; CHECK-NEXT:    [[TMP6:%.*]] = zext <4 x i1> [[TMP16]] to <4 x i32>
 ; CHECK-NEXT:    [[TMP12:%.*]] = shufflevector <4 x i32> [[TMP6]], <4 x i32> <i32 0, i32 0, i32 poison, i32 0>, <4 x i32> <i32 4, i32 5, i32 2, i32 7>
 ; CHECK-NEXT:    [[TMP13:%.*]] = select <4 x i1> [[TMP11]], <4 x i32> [[TMP12]], <4 x i32> zeroinitializer
 ; CHECK-NEXT:    [[TMP14:%.*]] = call i32 @llvm.vector.reduce.and.v4i32(<4 x i32> [[TMP13]])