[llvm] f676944 - [SLP][NFC] Use BoUpSLP::getNumberOfParts() where appropriate (#225995)

[via llvm-commits]

Author: Ryan Buchner
Date: 2026-09-24T17:50:39Z
New Revision: f676944065576ceba42691edeba053898b57c1bf

URL: https://github.com/llvm/llvm-project/commit/f676944065576ceba42691edeba053898b57c1bf
DIFF: https://github.com/llvm/llvm-project/commit/f676944065576ceba42691edeba053898b57c1bf.diff

LOG: [SLP][NFC] Use BoUpSLP::getNumberOfParts() where appropriate (#225995)

Caches the results for better performance rather than directly calling
slpvectorizer::getNumberOrParts().

Small improvement in compile time.

Added: 
    

Modified: 
    llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index f6c0f9e08a2af..6287cb856c2ed 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -6651,11 +6651,10 @@ BoUpSLP::getReorderingData(const TreeEntry &TE, bool TopToBottom,
       }
     }
     if (Sz == 2 && TE.getVectorFactor() == 4 &&
-        slpvectorizer::getNumberOfParts(
-            *TTI,
+        getNumberOfParts(
             getWidenedType(getValueType(TE.Scalars.front(), SLPReVec),
                            2 * TE.getVectorFactor()),
-            getValueType(TE.Scalars.front(), SLPReVec), SLPReVec) == 1)
+            getValueType(TE.Scalars.front(), SLPReVec)) == 1)
       return std::nullopt;
     if (TE.ReuseShuffleIndices.size() % Sz != 0)
       return std::nullopt;
@@ -10122,10 +10121,8 @@ static bool tryToFindDuplicates(SmallVectorImpl<Value *> &VL,
     auto *VecTy = cast<VectorType>(getWidenedType(ScalarTy, VL.size()));
     auto *UniquesVecTy =
         cast<VectorType>(getWidenedType(ScalarTy, NumUniqueScalarValues));
-    const unsigned NumParts =
-        slpvectorizer::getNumberOfParts(TTI, VecTy, ScalarTy, SLPReVec);
-    const unsigned UniquesNumParts =
-        slpvectorizer::getNumberOfParts(TTI, UniquesVecTy, ScalarTy, SLPReVec);
+    const unsigned NumParts = R.getNumberOfParts(VecTy, ScalarTy);
+    const unsigned UniquesNumParts = R.getNumberOfParts(UniquesVecTy, ScalarTy);
     // No need to schedule scalars and only single register used? Use original
     // scalars, do not pack.
     if (!RequireScheduling) {
@@ -14296,10 +14293,9 @@ void BoUpSLP::transformNodes() {
           bool IsSplat = isSplat(Slice);
           bool IsTwoRegisterSplat = true;
           if (IsSplat && VF == 2) {
-            unsigned NumRegs2VF = slpvectorizer::getNumberOfParts(
-                *TTI,
+            unsigned NumRegs2VF = getNumberOfParts(
                 getWidenedType(getValueType(Slice.front(), SLPReVec), 2 * VF),
-                getValueType(Slice.front(), SLPReVec), SLPReVec);
+                getValueType(Slice.front(), SLPReVec));
             IsTwoRegisterSplat = NumRegs2VF == 2;
           }
           if (Slices.empty() || !IsSplat || !IsTwoRegisterSplat ||
@@ -15540,8 +15536,7 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
     }
     assert(!CommonMask.empty() && "Expected non-empty common mask.");
     auto *MaskVecTy = getWidenedType(ScalarTy, Mask.size());
-    unsigned NumParts = slpvectorizer::getNumberOfParts(
-        TTI, MaskVecTy, ScalarTy, SLPReVec, Mask.size());
+    unsigned NumParts = R.getNumberOfParts(MaskVecTy, ScalarTy, Mask.size());
     unsigned SliceSize = getPartNumElems(Mask.size(), NumParts);
     const auto *It = find_if(Mask, not_equal_to(PoisonMaskElem));
     unsigned Part = std::distance(Mask.begin(), It) / SliceSize;
@@ -15556,8 +15551,7 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
     }
     assert(!CommonMask.empty() && "Expected non-empty common mask.");
     auto *MaskVecTy = getWidenedType(ScalarTy, Mask.size());
-    unsigned NumParts = slpvectorizer::getNumberOfParts(
-        TTI, MaskVecTy, ScalarTy, SLPReVec, Mask.size());
+    unsigned NumParts = R.getNumberOfParts(MaskVecTy, ScalarTy, Mask.size());
     unsigned SliceSize = getPartNumElems(Mask.size(), NumParts);
     const auto *It = find_if(Mask, not_equal_to(PoisonMaskElem));
     unsigned Part = std::distance(Mask.begin(), It) / SliceSize;
@@ -26482,9 +26476,8 @@ void BoUpSLP::optimizeGatherSequence() {
   // and its mask indeces are the same as in the first one or undefs. E.g.
   // shuffle %0, poison, <0, 0, 0, undef> is less defined than shuffle %0,
   // poison, <0, 0, 0, 0>.
-  auto &&IsIdenticalOrLessDefined = [TTI = TTI](Instruction *I1,
-                                                Instruction *I2,
-                                                SmallVectorImpl<int> &NewMask) {
+  auto &&IsIdenticalOrLessDefined = [this](Instruction *I1, Instruction *I2,
+                                           SmallVectorImpl<int> &NewMask) {
     if (I1->getType() != I2->getType())
       return false;
     auto *SI1 = dyn_cast<ShuffleVectorInst>(I1);
@@ -26516,14 +26509,10 @@ void BoUpSLP::optimizeGatherSequence() {
     // Check if the last undefs actually change the final number of used vector
     // registers.
     return SM1.size() - LastUndefsCnt > 1 &&
-           slpvectorizer::getNumberOfParts(*TTI, SI1->getType(),
-                                           SI1->getType()->getElementType(),
-                                           SLPReVec) ==
-               slpvectorizer::getNumberOfParts(
-                   *TTI,
-                   getWidenedType(SI1->getType()->getElementType(),
-                                  SM1.size() - LastUndefsCnt),
-                   SI1->getType()->getElementType(), SLPReVec);
+           getNumberOfParts(SI1->getType(), SI1->getType()->getElementType()) ==
+               getNumberOfParts(getWidenedType(SI1->getType()->getElementType(),
+                                               SM1.size() - LastUndefsCnt),
+                                SI1->getType()->getElementType());
   };
   // Perform O(N^2) search over the gather/shuffle sequences and merge identical
   // instructions. TODO: We can further optimize this scan if we split the
@@ -32375,7 +32364,7 @@ class HorizontalReduction {
         ReduxWidth = getFloorFullVectorNumberOfElements(TTI, ScalarTy,
                                                         ReduxWidth, SLPReVec);
         VectorType *Tp = cast<VectorType>(getWidenedType(ScalarTy, ReduxWidth));
-        NumParts = slpvectorizer::getNumberOfParts(TTI, Tp, ScalarTy, SLPReVec);
+        NumParts = V.getNumberOfParts(Tp, ScalarTy);
         NumRegs =
             TTI.getNumberOfRegisters(TTI.getRegisterClassForType(true, Tp));
         while (NumParts > NumRegs) {
@@ -32383,8 +32372,7 @@ class HorizontalReduction {
           ReduxWidth = bit_floor(ReduxWidth - 1);
           VectorType *Tp =
               cast<VectorType>(getWidenedType(ScalarTy, ReduxWidth));
-          NumParts =
-              slpvectorizer::getNumberOfParts(TTI, Tp, ScalarTy, SLPReVec);
+          NumParts = V.getNumberOfParts(Tp, ScalarTy);
           NumRegs =
               TTI.getNumberOfRegisters(TTI.getRegisterClassForType(true, Tp));
         }
@@ -33091,15 +33079,14 @@ class HorizontalReduction {
       ReduxWidth = getFloorFullVectorNumberOfElements(TTI, ScalarTy, ReduxWidth,
                                                       SLPReVec);
       Type *Tp = getWidenedType(ScalarTy, ReduxWidth);
-      unsigned NumParts =
-          slpvectorizer::getNumberOfParts(TTI, Tp, ScalarTy, SLPReVec);
+      unsigned NumParts = V.getNumberOfParts(Tp, ScalarTy);
       unsigned NumRegs =
           TTI.getNumberOfRegisters(TTI.getRegisterClassForType(true, Tp));
       while (NumParts > NumRegs) {
         assert(ReduxWidth > 0 && "ReduxWidth is unexpectedly 0.");
         ReduxWidth = bit_floor(ReduxWidth - 1);
         Type *Tp = getWidenedType(ScalarTy, ReduxWidth);
-        NumParts = slpvectorizer::getNumberOfParts(TTI, Tp, ScalarTy, SLPReVec);
+        NumParts = V.getNumberOfParts(Tp, ScalarTy);
         NumRegs =
             TTI.getNumberOfRegisters(TTI.getRegisterClassForType(true, Tp));
       }