[clang] [llvm] [VectorCombine] Shrink loads used in shufflevector rebroadcasts (PR #128938)
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Wed Jun 11 01:23:56 PDT 2025
================
@@ -3483,6 +3487,125 @@ 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 *OldLoadTy = dyn_cast<FixedVectorType>(OldLoad->getType());
+ if (!OldLoadTy)
+ return false;
+
+ unsigned const OldNumElements = OldLoadTy->getNumElements();
+
+ // 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(OldNumElements, -1);
+ for (auto &Use : I.uses()) {
+ // Ensure all uses match the required pattern.
+ User *Shuffle = Use.getUser();
+ ArrayRef<int> Mask;
+
+ if (!match(Shuffle,
+ m_Shuffle(m_Specific(OldLoad), 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.
+ for (int Index : Mask) {
+ if (Index >= 0 && Index < static_cast<int>(OldNumElements)) {
+ 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 const NewNumElements = Indices->second + 1u;
+
+ // If the range of vector elements is smaller than the full load, attempt
+ // to create a smaller load.
+ if (NewNumElements < OldNumElements) {
+ auto Builder = IRBuilder(&I);
+ Builder.SetCurrentDebugLocation(I.getDebugLoc());
+
+ // Calculate costs of old and new ops.
+ Type *ElemTy = OldLoadTy->getElementType();
+ FixedVectorType *NewLoadTy = FixedVectorType::get(ElemTy, NewNumElements);
+ Value *PtrOp = OldLoad->getPointerOperand();
+
+ InstructionCost OldCost = TTI.getMemoryOpCost(
+ Instruction::Load, OldLoad->getType(), OldLoad->getAlign(),
+ OldLoad->getPointerAddressSpace(), CostKind);
+ InstructionCost NewCost =
+ TTI.getMemoryOpCost(Instruction::Load, NewLoadTy, OldLoad->getAlign(),
+ OldLoad->getPointerAddressSpace(), CostKind);
+
+ using UseEntry = std::pair<ShuffleVectorInst *, std::vector<int>>;
+ auto NewUses = SmallVector<UseEntry, 4u>();
+ auto SizeDiff = OldNumElements - NewNumElements;
----------------
RKSimon wrote:
(style) don't use auto except for casts/iterators
https://github.com/llvm/llvm-project/pull/128938
More information about the llvm-commits
mailing list