From b748675cd8dfe5f9c23aa83c19f58b5261dcc0c8 Mon Sep 17 00:00:00 2001 From: Philip Reames Date: Thu, 24 Jul 2025 13:53:00 -0700 Subject: [PATCH] [RISCV][IA] Rework VL for strided LD/ST optimization [nfc] I'd originally written this creating a new VL, but realized it was probably cleaner to be explicit about the truncation which is happening. As these VLs are all constants (since these two codepaths see fixed vectors) this should result in identical constants being created. --- llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp | 15 +++++++-------- 1 file changed, 7 insertions(+), 8 deletions(-) diff --git a/llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp b/llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp index 0565fcd9c6bcf..30d8f850763a2 100644 --- a/llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp +++ b/llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp @@ -224,10 +224,10 @@ bool RISCVTargetLowering::lowerInterleavedLoad( Value *Stride = ConstantInt::get(XLenTy, Factor * ScalarSizeInBytes); Value *Offset = ConstantInt::get(XLenTy, Indices[0] * ScalarSizeInBytes); Value *BasePtr = Builder.CreatePtrAdd(Ptr, Offset); - // Note: Same VL as above, but i32 not xlen due to signature of - // vp.strided.load - VL = Builder.CreateElementCount(Builder.getInt32Ty(), - VTy->getElementCount()); + // For rv64, need to truncate i64 to i32 to match signature. As VL is at most + // the number of active lanes (which is bounded by i32) this is safe. + VL = Builder.CreateTrunc(VL, Builder.getInt32Ty()); + CallInst *CI = Builder.CreateIntrinsic(Intrinsic::experimental_vp_strided_load, {VTy, BasePtr->getType(), Stride->getType()}, @@ -302,10 +302,9 @@ bool RISCVTargetLowering::lowerInterleavedStore(Instruction *Store, Value *Stride = ConstantInt::get(XLenTy, Factor * ScalarSizeInBytes); Value *Offset = ConstantInt::get(XLenTy, Index * ScalarSizeInBytes); Value *BasePtr = Builder.CreatePtrAdd(Ptr, Offset); - // Note: Same VL as above, but i32 not xlen due to signature of - // vp.strided.store - VL = Builder.CreateElementCount(Builder.getInt32Ty(), - VTy->getElementCount()); + // For rv64, need to truncate i64 to i32 to match signature. As VL is at + // most the number of active lanes (which is bounded by i32) this is safe. + VL = Builder.CreateTrunc(VL, Builder.getInt32Ty()); CallInst *CI = Builder.CreateIntrinsic(Intrinsic::experimental_vp_strided_store,