[clang] [llvm] [VectorCombine] Shrink loads used in shufflevector rebroadcasts (PR #128938)
Simon Pilgrim via cfe-commits
cfe-commits at lists.llvm.org
Thu May 29 01:39:46 PDT 2025
================
@@ -3483,6 +3487,127 @@ bool VectorCombine::foldInterleaveIntrinsics(Instruction &I) {
return true;
}
+// Attempt to shrink loads that are only used by shufflevector instructions.
+bool VectorCombine::shrinkLoadForShuffles(Instruction &I) {
+ auto *OldLoad = dyn_cast<LoadInst>(&I);
+ if (!OldLoad || !OldLoad->isSimple())
+ return false;
+
+ auto *VecTy = dyn_cast<FixedVectorType>(OldLoad->getType());
+ if (!VecTy)
+ return false;
+
+ // Search all uses of load. If all uses are shufflevector instructions, and
+ // the second operands are all poison values, find the minimum and maximum
+ // indices of the vector elements referenced by all shuffle masks.
+ // Otherwise return `std::nullopt`.
+ using IndexRange = std::pair<int, int>;
+ auto GetIndexRangeInShuffles = [&]() -> std::optional<IndexRange> {
+ IndexRange OutputRange = IndexRange(VecTy->getNumElements(), -1);
+ for (auto &Use : I.uses()) {
+ // Ensure all uses match the required pattern.
+ User *Shuffle = Use.getUser();
+ Value *Op0 = nullptr;
+ ArrayRef<int> Mask;
+
+ if (!match(Shuffle, m_Shuffle(m_Value(Op0), m_Undef(), m_Mask(Mask))))
+ return std::nullopt;
+
+ // Ignore shufflevector instructions that have no uses.
+ if (Shuffle->use_empty())
+ continue;
+
+ // Find the min and max indices used by the shufflevector instruction.
+ FixedVectorType *Op0Ty = cast<FixedVectorType>(Op0->getType());
+ int NumElems = static_cast<int>(Op0Ty->getNumElements());
+
+ for (int Index : Mask) {
+ if (Index >= 0) {
+ Index %= NumElems;
+ OutputRange.first = std::min(Index, OutputRange.first);
+ OutputRange.second = std::max(Index, OutputRange.second);
+ }
+ }
+ }
+
+ if (OutputRange.second < OutputRange.first)
+ return std::nullopt;
+
+ return OutputRange;
+ };
+
+ // Get the range of vector elements used by shufflevector instructions.
+ if (auto Indices = GetIndexRangeInShuffles()) {
+ unsigned OldSize = VecTy->getNumElements();
+ unsigned NewSize = Indices->second + 1u;
+
+ // If the range of vector elements is smaller than the full load, attempt
+ // to create a smaller load.
+ if (NewSize < OldSize) {
+ auto Builder = IRBuilder(&I);
+ Builder.SetCurrentDebugLocation(I.getDebugLoc());
+
+ // Create new load of smaller vector.
+ auto *ElemTy = VecTy->getElementType();
+ auto *NewVecTy = FixedVectorType::get(ElemTy, NewSize);
+ auto *PtrOp = OldLoad->getPointerOperand();
+ auto *NewLoad = cast<LoadInst>(
+ Builder.CreateAlignedLoad(NewVecTy, PtrOp, OldLoad->getAlign()));
+ NewLoad->copyMetadata(I);
+
+ // Calculate costs of old and new ops.
+ auto OldCost = TTI.getMemoryOpCost(
+ Instruction::Load, OldLoad->getType(), OldLoad->getAlign(),
+ OldLoad->getPointerAddressSpace(), CostKind);
+ auto NewCost = TTI.getMemoryOpCost(
----------------
RKSimon wrote:
(style) avoid auto in most cases
https://github.com/llvm/llvm-project/pull/128938
More information about the cfe-commits
mailing list