From 5da23f2de29116cb422c011279eea89011419a20 Mon Sep 17 00:00:00 2001 From: Damian Heaton Date: Fri, 19 Sep 2025 10:51:26 +0000 Subject: [PATCH 1/7] Add `llvm.vector.partial.reduction.fadd` intrinsic With this intrinsic, and supporting SelectionDAG nodes, we can better make use of instructions such as AArch64's `FDOT`. --- llvm/docs/LangRef.rst | 42 ++++++++++++ llvm/include/llvm/CodeGen/ISDOpcodes.h | 3 +- llvm/include/llvm/CodeGen/TargetLowering.h | 4 +- llvm/include/llvm/IR/Intrinsics.td | 4 ++ .../include/llvm/Target/TargetSelectionDAG.td | 2 + llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 22 +++++-- .../SelectionDAG/LegalizeVectorOps.cpp | 2 + .../SelectionDAG/LegalizeVectorTypes.cpp | 2 + .../lib/CodeGen/SelectionDAG/SelectionDAG.cpp | 3 +- .../SelectionDAG/SelectionDAGBuilder.cpp | 13 ++++ .../SelectionDAG/SelectionDAGDumper.cpp | 2 + .../CodeGen/SelectionDAG/TargetLowering.cpp | 21 +++--- .../Target/AArch64/AArch64ISelLowering.cpp | 6 ++ .../lib/Target/AArch64/AArch64SVEInstrInfo.td | 3 + llvm/test/CodeGen/AArch64/sve2p1-fdot.ll | 66 +++++++++++++++++++ 15 files changed, 177 insertions(+), 18 deletions(-) create mode 100644 llvm/test/CodeGen/AArch64/sve2p1-fdot.ll diff --git a/llvm/docs/LangRef.rst b/llvm/docs/LangRef.rst index 8e863939781a2..d89201477e4b1 100644 --- a/llvm/docs/LangRef.rst +++ b/llvm/docs/LangRef.rst @@ -20732,6 +20732,48 @@ performance, and an out-of-loop phase to calculate the final scalar result. By avoiding the introduction of new ordering constraints, these intrinsics enhance the ability to leverage a target's accumulation instructions. +'``llvm.vector.partial.reduce.fadd.*``' Intrinsic +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + +Syntax: +""""""" +This is an overloaded intrinsic. + +:: + + declare <4 x f32> @llvm.vector.partial.reduce.fadd.v4f32.v4f32.v8f32(<4 x f32> %a, <8 x f32> %b) + declare @llvm.vector.partial.reduce.add.nxv4f32.nxv4f32.nxv8f32( %a, %b) + +Overview: +""""""""" + +The '``llvm.vector.partial.reduce.fadd.*``' intrinsics reduce the +concatenation of the two vector arguments down to the number of elements of the +result vector type. + +Arguments: +"""""""""" + +The first argument is a floating-point vector with the same type as the result. + +The second argument is a vector with a length that is a known integer multiple +of the result's type, while maintaining the same element type. + +Semantics: +"""""""""" + +Other than the reduction operator (e.g. add) the way in which the concatenated +arguments is reduced is entirely unspecified. By their nature these intrinsics +are not expected to be useful in isolation but instead implement the first phase +of an overall reduction operation. + +The typical use case is loop vectorization where reductions are split into an +in-loop phase, where maintaining an unordered vector result is important for +performance, and an out-of-loop phase to calculate the final scalar result. + +By avoiding the introduction of new ordering constraints, these intrinsics +enhance the ability to leverage a target's accumulation instructions. + '``llvm.experimental.vector.histogram.*``' Intrinsic ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ diff --git a/llvm/include/llvm/CodeGen/ISDOpcodes.h b/llvm/include/llvm/CodeGen/ISDOpcodes.h index c76c83d84b3c7..83ee6ff677e3d 100644 --- a/llvm/include/llvm/CodeGen/ISDOpcodes.h +++ b/llvm/include/llvm/CodeGen/ISDOpcodes.h @@ -1510,6 +1510,7 @@ enum NodeType { PARTIAL_REDUCE_SMLA, // sext, sext PARTIAL_REDUCE_UMLA, // zext, zext PARTIAL_REDUCE_SUMLA, // sext, zext + PARTIAL_REDUCE_FMLA, // fpext, fpext // The `llvm.experimental.stackmap` intrinsic. // Operands: input chain, glue, , , [live0[, live1...]] @@ -1761,7 +1762,7 @@ LLVM_ABI CondCode getSetCCInverse(CondCode Operation, EVT Type); inline bool isExtOpcode(unsigned Opcode) { return Opcode == ISD::ANY_EXTEND || Opcode == ISD::ZERO_EXTEND || - Opcode == ISD::SIGN_EXTEND; + Opcode == ISD::SIGN_EXTEND || Opcode == ISD::FP_EXTEND; } inline bool isExtVecInRegOpcode(unsigned Opcode) { diff --git a/llvm/include/llvm/CodeGen/TargetLowering.h b/llvm/include/llvm/CodeGen/TargetLowering.h index c45e03a7bdad8..8d75b0581618c 100644 --- a/llvm/include/llvm/CodeGen/TargetLowering.h +++ b/llvm/include/llvm/CodeGen/TargetLowering.h @@ -1672,7 +1672,7 @@ class LLVM_ABI TargetLoweringBase { LegalizeAction getPartialReduceMLAAction(unsigned Opc, EVT AccVT, EVT InputVT) const { assert(Opc == ISD::PARTIAL_REDUCE_SMLA || Opc == ISD::PARTIAL_REDUCE_UMLA || - Opc == ISD::PARTIAL_REDUCE_SUMLA); + Opc == ISD::PARTIAL_REDUCE_SUMLA || Opc == ISD::PARTIAL_REDUCE_FMLA); PartialReduceActionTypes Key = {Opc, AccVT.getSimpleVT().SimpleTy, InputVT.getSimpleVT().SimpleTy}; auto It = PartialReduceMLAActions.find(Key); @@ -2774,7 +2774,7 @@ class LLVM_ABI TargetLoweringBase { void setPartialReduceMLAAction(unsigned Opc, MVT AccVT, MVT InputVT, LegalizeAction Action) { assert(Opc == ISD::PARTIAL_REDUCE_SMLA || Opc == ISD::PARTIAL_REDUCE_UMLA || - Opc == ISD::PARTIAL_REDUCE_SUMLA); + Opc == ISD::PARTIAL_REDUCE_SUMLA || Opc == ISD::PARTIAL_REDUCE_FMLA); assert(AccVT.isValid() && InputVT.isValid() && "setPartialReduceMLAAction types aren't valid"); PartialReduceActionTypes Key = {Opc, AccVT.SimpleTy, InputVT.SimpleTy}; diff --git a/llvm/include/llvm/IR/Intrinsics.td b/llvm/include/llvm/IR/Intrinsics.td index dba44e7c3c506..9ef3705250b46 100644 --- a/llvm/include/llvm/IR/Intrinsics.td +++ b/llvm/include/llvm/IR/Intrinsics.td @@ -2802,6 +2802,10 @@ def int_vector_partial_reduce_add : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyvector_ty, llvm_anyvector_ty], [IntrNoMem]>; +def int_vector_partial_reduce_fadd : DefaultAttrsIntrinsic<[LLVMMatchType<0>], + [llvm_anyfloat_ty, llvm_anyfloat_ty], + [IntrNoMem]>; + //===----------------- Pointer Authentication Intrinsics ------------------===// // diff --git a/llvm/include/llvm/Target/TargetSelectionDAG.td b/llvm/include/llvm/Target/TargetSelectionDAG.td index 5e57dcaa303f3..1f0800885937e 100644 --- a/llvm/include/llvm/Target/TargetSelectionDAG.td +++ b/llvm/include/llvm/Target/TargetSelectionDAG.td @@ -522,6 +522,8 @@ def partial_reduce_smla : SDNode<"ISD::PARTIAL_REDUCE_SMLA", SDTPartialReduceMLA>; def partial_reduce_sumla : SDNode<"ISD::PARTIAL_REDUCE_SUMLA", SDTPartialReduceMLA>; +def partial_reduce_fmla : SDNode<"ISD::PARTIAL_REDUCE_FMLA", + SDTPartialReduceMLA>; def fadd : SDNode<"ISD::FADD" , SDTFPBinOp, [SDNPCommutative]>; def fsub : SDNode<"ISD::FSUB" , SDTFPBinOp>; diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp index c81568672de3c..8615d3060b8e8 100644 --- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -2040,6 +2040,7 @@ SDValue DAGCombiner::visit(SDNode *N) { case ISD::PARTIAL_REDUCE_SMLA: case ISD::PARTIAL_REDUCE_UMLA: case ISD::PARTIAL_REDUCE_SUMLA: + case ISD::PARTIAL_REDUCE_FMLA: return visitPARTIAL_REDUCE_MLA(N); case ISD::VECTOR_COMPRESS: return visitVECTOR_COMPRESS(N); case ISD::LIFETIME_END: return visitLIFETIME_END(N); @@ -12997,8 +12998,11 @@ SDValue DAGCombiner::foldPartialReduceMLAMulOp(SDNode *N) { SDValue Op2 = N->getOperand(2); APInt C; - if (Op1->getOpcode() != ISD::MUL || - !ISD::isConstantSplatVector(Op2.getNode(), C) || !C.isOne()) + if (!(Op1->getOpcode() == ISD::MUL && + ISD::isConstantSplatVector(Op2.getNode(), C) && C.isOne()) && + !(Op1->getOpcode() == ISD::FMUL && + ISD::isConstantSplatVector(Op2.getNode(), C) && + C == APFloat(1.0f).bitcastToAPInt().trunc(C.getBitWidth()))) return SDValue(); SDValue LHS = Op1->getOperand(0); @@ -13053,6 +13057,8 @@ SDValue DAGCombiner::foldPartialReduceMLAMulOp(SDNode *N) { else if (LHSOpcode == ISD::ZERO_EXTEND && RHSOpcode == ISD::SIGN_EXTEND) { NewOpc = ISD::PARTIAL_REDUCE_SUMLA; std::swap(LHSExtOp, RHSExtOp); + } else if (LHSOpcode == ISD::FP_EXTEND && RHSOpcode == ISD::FP_EXTEND) { + NewOpc = ISD::PARTIAL_REDUCE_FMLA; } else return SDValue(); // For a 2-stage extend the signedness of both of the extends must match @@ -13088,22 +13094,26 @@ SDValue DAGCombiner::foldPartialReduceAdd(SDNode *N) { APInt ConstantOne; if (!ISD::isConstantSplatVector(Op2.getNode(), ConstantOne) || - !ConstantOne.isOne()) + !(ConstantOne.isOne() || + ConstantOne == + APFloat(1.0f).bitcastToAPInt().trunc(ConstantOne.getBitWidth()))) return SDValue(); unsigned Op1Opcode = Op1.getOpcode(); if (!ISD::isExtOpcode(Op1Opcode)) return SDValue(); - bool Op1IsSigned = Op1Opcode == ISD::SIGN_EXTEND; + bool Op1IsSigned = Op1Opcode != ISD::ZERO_EXTEND; bool NodeIsSigned = N->getOpcode() != ISD::PARTIAL_REDUCE_UMLA; EVT AccElemVT = Acc.getValueType().getVectorElementType(); if (Op1IsSigned != NodeIsSigned && Op1.getValueType().getVectorElementType() != AccElemVT) return SDValue(); - unsigned NewOpcode = - Op1IsSigned ? ISD::PARTIAL_REDUCE_SMLA : ISD::PARTIAL_REDUCE_UMLA; + unsigned NewOpcode = N->getOpcode() == ISD::PARTIAL_REDUCE_FMLA + ? ISD::PARTIAL_REDUCE_FMLA + : Op1IsSigned ? ISD::PARTIAL_REDUCE_SMLA + : ISD::PARTIAL_REDUCE_UMLA; SDValue UnextOp1 = Op1.getOperand(0); EVT UnextOp1VT = UnextOp1.getValueType(); diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp index 8e423c4f83b38..94751be5b7986 100644 --- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp @@ -534,6 +534,7 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) { case ISD::PARTIAL_REDUCE_UMLA: case ISD::PARTIAL_REDUCE_SMLA: case ISD::PARTIAL_REDUCE_SUMLA: + case ISD::PARTIAL_REDUCE_FMLA: Action = TLI.getPartialReduceMLAAction(Op.getOpcode(), Node->getValueType(0), Node->getOperand(1).getValueType()); @@ -1243,6 +1244,7 @@ void VectorLegalizer::Expand(SDNode *Node, SmallVectorImpl &Results) { case ISD::PARTIAL_REDUCE_UMLA: case ISD::PARTIAL_REDUCE_SMLA: case ISD::PARTIAL_REDUCE_SUMLA: + case ISD::PARTIAL_REDUCE_FMLA: Results.push_back(TLI.expandPartialReduceMLA(Node, DAG)); return; case ISD::VECREDUCE_SEQ_FADD: diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp index ff7cd665446cc..e6f19499b8f41 100644 --- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp @@ -1459,6 +1459,7 @@ void DAGTypeLegalizer::SplitVectorResult(SDNode *N, unsigned ResNo) { case ISD::PARTIAL_REDUCE_UMLA: case ISD::PARTIAL_REDUCE_SMLA: case ISD::PARTIAL_REDUCE_SUMLA: + case ISD::PARTIAL_REDUCE_FMLA: SplitVecRes_PARTIAL_REDUCE_MLA(N, Lo, Hi); break; case ISD::GET_ACTIVE_LANE_MASK: @@ -3674,6 +3675,7 @@ bool DAGTypeLegalizer::SplitVectorOperand(SDNode *N, unsigned OpNo) { case ISD::PARTIAL_REDUCE_UMLA: case ISD::PARTIAL_REDUCE_SMLA: case ISD::PARTIAL_REDUCE_SUMLA: + case ISD::PARTIAL_REDUCE_FMLA: Res = SplitVecOp_PARTIAL_REDUCE_MLA(N); break; } diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp index 7aa293af963e6..0ccaf9c7d887d 100644 --- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp @@ -8394,7 +8394,8 @@ SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT, } case ISD::PARTIAL_REDUCE_UMLA: case ISD::PARTIAL_REDUCE_SMLA: - case ISD::PARTIAL_REDUCE_SUMLA: { + case ISD::PARTIAL_REDUCE_SUMLA: + case ISD::PARTIAL_REDUCE_FMLA: { [[maybe_unused]] EVT AccVT = N1.getValueType(); [[maybe_unused]] EVT Input1VT = N2.getValueType(); [[maybe_unused]] EVT Input2VT = N3.getValueType(); diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp index b5201a311c591..544b4abd6b45a 100644 --- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp @@ -8114,6 +8114,19 @@ void SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, Input, DAG.getConstant(1, sdl, Input.getValueType()))); return; } + case Intrinsic::vector_partial_reduce_fadd: { + if (!TLI.shouldExpandPartialReductionIntrinsic(cast(&I))) { + visitTargetIntrinsic(I, Intrinsic); + return; + } + SDValue Acc = getValue(I.getOperand(0)); + SDValue Input = getValue(I.getOperand(1)); + setValue(&I, + DAG.getNode(ISD::PARTIAL_REDUCE_FMLA, sdl, Acc.getValueType(), Acc, + Input, + DAG.getConstantFP(1.0f, sdl, Input.getValueType()))); + return; + } case Intrinsic::experimental_cttz_elts: { auto DL = getCurSDLoc(); SDValue Op = getValue(I.getOperand(0)); diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp index 4b2a00c2e2cfa..cf5c269c20761 100644 --- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp @@ -587,6 +587,8 @@ std::string SDNode::getOperationName(const SelectionDAG *G) const { return "partial_reduce_smla"; case ISD::PARTIAL_REDUCE_SUMLA: return "partial_reduce_sumla"; + case ISD::PARTIAL_REDUCE_FMLA: + return "partial_reduce_fmla"; case ISD::LOOP_DEPENDENCE_WAR_MASK: return "loop_dep_war"; case ISD::LOOP_DEPENDENCE_RAW_MASK: diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp index dba5a8c0a7315..8e284997af01c 100644 --- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp @@ -12062,12 +12062,14 @@ SDValue TargetLowering::expandPartialReduceMLA(SDNode *N, EVT::getVectorVT(*DAG.getContext(), AccVT.getVectorElementType(), MulOpVT.getVectorElementCount()); - unsigned ExtOpcLHS = N->getOpcode() == ISD::PARTIAL_REDUCE_UMLA - ? ISD::ZERO_EXTEND - : ISD::SIGN_EXTEND; - unsigned ExtOpcRHS = N->getOpcode() == ISD::PARTIAL_REDUCE_SMLA - ? ISD::SIGN_EXTEND - : ISD::ZERO_EXTEND; + unsigned ExtOpcLHS = + N->getOpcode() == ISD::PARTIAL_REDUCE_FMLA ? ISD::FP_EXTEND + : N->getOpcode() == ISD::PARTIAL_REDUCE_UMLA ? ISD::ZERO_EXTEND + : ISD::SIGN_EXTEND; + unsigned ExtOpcRHS = + N->getOpcode() == ISD::PARTIAL_REDUCE_FMLA ? ISD::FP_EXTEND + : N->getOpcode() == ISD::PARTIAL_REDUCE_SMLA ? ISD::SIGN_EXTEND + : ISD::ZERO_EXTEND; if (ExtMulOpVT != MulOpVT) { MulLHS = DAG.getNode(ExtOpcLHS, DL, ExtMulOpVT, MulLHS); @@ -12076,7 +12078,7 @@ SDValue TargetLowering::expandPartialReduceMLA(SDNode *N, SDValue Input = MulLHS; APInt ConstantOne; if (!ISD::isConstantSplatVector(MulRHS.getNode(), ConstantOne) || - !ConstantOne.isOne()) + !(ConstantOne.isOne() || ConstantOne == APFloat(1.0f).bitcastToAPInt())) Input = DAG.getNode(ISD::MUL, DL, ExtMulOpVT, MulLHS, MulRHS); unsigned Stride = AccVT.getVectorMinNumElements(); @@ -12087,10 +12089,13 @@ SDValue TargetLowering::expandPartialReduceMLA(SDNode *N, for (unsigned I = 0; I < ScaleFactor; I++) Subvectors.push_back(DAG.getExtractSubvector(DL, AccVT, Input, I * Stride)); + unsigned FlatNode = + N->getOpcode() == ISD::PARTIAL_REDUCE_FMLA ? ISD::FADD : ISD::ADD; + // Flatten the subvector tree while (Subvectors.size() > 1) { Subvectors.push_back( - DAG.getNode(ISD::ADD, DL, AccVT, {Subvectors[0], Subvectors[1]})); + DAG.getNode(FlatNode, DL, AccVT, {Subvectors[0], Subvectors[1]})); Subvectors.pop_front(); Subvectors.pop_front(); } diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp index a4c1e265f0e63..2733e46cf94bb 100644 --- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp +++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp @@ -1919,6 +1919,12 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM, } } + // Handle floating-point partial reduction + if (Subtarget->hasSVE2p1() || Subtarget->hasSME2()) { + static const unsigned FMLAOps[] = {ISD::PARTIAL_REDUCE_FMLA}; + setPartialReduceMLAAction(FMLAOps, MVT::nxv4f32, MVT::nxv8f16, Legal); + } + // Handle non-aliasing elements mask if (Subtarget->hasSVE2() || (Subtarget->hasSME() && Subtarget->isStreaming())) { diff --git a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td index 1e30735b7a56a..9721641d8bde0 100644 --- a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td +++ b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td @@ -4228,6 +4228,9 @@ defm FCLAMP_ZZZ : sve_fp_clamp<"fclamp", AArch64fclamp>; defm FDOT_ZZZ_S : sve_float_dot<0b0, 0b0, ZPR32, ZPR16, "fdot", nxv8f16, int_aarch64_sve_fdot_x2>; defm FDOT_ZZZI_S : sve_float_dot_indexed<0b0, 0b00, ZPR16, ZPR3b16, "fdot", nxv8f16, int_aarch64_sve_fdot_lane_x2>; +def : Pat<(nxv4f32 (partial_reduce_fmla nxv4f32:$Acc, nxv8f16:$LHS, nxv8f16:$RHS)), + (FDOT_ZZZ_S $Acc, $LHS, $RHS)>; + defm BFMLSLB_ZZZ_S : sve2_fp_mla_long<0b110, "bfmlslb", nxv4f32, nxv8bf16, int_aarch64_sve_bfmlslb>; defm BFMLSLT_ZZZ_S : sve2_fp_mla_long<0b111, "bfmlslt", nxv4f32, nxv8bf16, int_aarch64_sve_bfmlslt>; defm BFMLSLB_ZZZI_S : sve2_fp_mla_long_by_indexed_elem<0b110, "bfmlslb", nxv4f32, nxv8bf16, int_aarch64_sve_bfmlslb_lane>; diff --git a/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll b/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll new file mode 100644 index 0000000000000..5bb1fae43392f --- /dev/null +++ b/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll @@ -0,0 +1,66 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 +; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve2p1 < %s | FileCheck %s + +define @fdot_wide_vl128( %acc, %a, %b) { +; CHECK-LABEL: fdot_wide_vl128: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: fdot z0.s, z1.h, z2.h +; CHECK-NEXT: ret +entry: + %a.wide = fpext %a to + %b.wide = fpext %b to + %mult = fmul %a.wide, %b.wide + %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %mult) + ret %partial.reduce +} + +define void @fdot_wide_vl256(ptr %accptr, ptr %aptr, ptr %bptr) vscale_range(2,2) { +; CHECK-LABEL: fdot_wide_vl256: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: ptrue p0.s +; CHECK-NEXT: ld1h { z0.s }, p0/z, [x1] +; CHECK-NEXT: ld1h { z1.s }, p0/z, [x2] +; CHECK-NEXT: ld1h { z2.s }, p0/z, [x1, #1, mul vl] +; CHECK-NEXT: ld1h { z3.s }, p0/z, [x2, #1, mul vl] +; CHECK-NEXT: fcvt z0.s, p0/m, z0.h +; CHECK-NEXT: fcvt z1.s, p0/m, z1.h +; CHECK-NEXT: fcvt z2.s, p0/m, z2.h +; CHECK-NEXT: fcvt z3.s, p0/m, z3.h +; CHECK-NEXT: fmul z0.s, z0.s, z1.s +; CHECK-NEXT: ldr z1, [x0] +; CHECK-NEXT: fmul z2.s, z2.s, z3.s +; CHECK-NEXT: fadd z0.s, z1.s, z0.s +; CHECK-NEXT: fadd z0.s, z0.s, z2.s +; CHECK-NEXT: str z0, [x0] +; CHECK-NEXT: ret +entry: + %acc = load <8 x float>, ptr %accptr + %a = load <16 x half>, ptr %aptr + %b = load <16 x half>, ptr %bptr + %a.wide = fpext <16 x half> %a to <16 x float> + %b.wide = fpext <16 x half> %b to <16 x float> + %mult = fmul <16 x float> %a.wide, %b.wide + %partial.reduce = call <8 x float> @llvm.vector.partial.reduce.fadd(<8 x float> %acc, <16 x float> %mult) + store <8 x float> %partial.reduce, ptr %accptr + ret void +} + +define <4 x float> @fixed_fdot_wide(<4 x float> %acc, <8 x half> %a, <8 x half> %b) { +; CHECK-LABEL: fixed_fdot_wide: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: fcvtl v3.4s, v1.4h +; CHECK-NEXT: fcvtl v4.4s, v2.4h +; CHECK-NEXT: fcvtl2 v1.4s, v1.8h +; CHECK-NEXT: fcvtl2 v2.4s, v2.8h +; CHECK-NEXT: fmul v3.4s, v3.4s, v4.4s +; CHECK-NEXT: fmul v1.4s, v1.4s, v2.4s +; CHECK-NEXT: fadd v0.4s, v0.4s, v3.4s +; CHECK-NEXT: fadd v0.4s, v0.4s, v1.4s +; CHECK-NEXT: ret +entry: + %a.wide = fpext <8 x half> %a to <8 x float> + %b.wide = fpext <8 x half> %b to <8 x float> + %mult = fmul <8 x float> %a.wide, %b.wide + %partial.reduce = call <4 x float> @llvm.vector.partial.reduce.fadd(<4 x float> %acc, <8 x float> %mult) + ret <4 x float> %partial.reduce +} From 9d452270e2796a1dc82ea2e2bc3606ba859af54d Mon Sep 17 00:00:00 2001 From: Damian Heaton Date: Fri, 19 Sep 2025 16:00:49 +0000 Subject: [PATCH 2/7] Revert adding `FP_EXTEND` to `isExtOpcode` --- llvm/include/llvm/CodeGen/ISDOpcodes.h | 2 +- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 4 ++-- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/llvm/include/llvm/CodeGen/ISDOpcodes.h b/llvm/include/llvm/CodeGen/ISDOpcodes.h index 83ee6ff677e3d..e1f6aab0040db 100644 --- a/llvm/include/llvm/CodeGen/ISDOpcodes.h +++ b/llvm/include/llvm/CodeGen/ISDOpcodes.h @@ -1762,7 +1762,7 @@ LLVM_ABI CondCode getSetCCInverse(CondCode Operation, EVT Type); inline bool isExtOpcode(unsigned Opcode) { return Opcode == ISD::ANY_EXTEND || Opcode == ISD::ZERO_EXTEND || - Opcode == ISD::SIGN_EXTEND || Opcode == ISD::FP_EXTEND; + Opcode == ISD::SIGN_EXTEND; } inline bool isExtVecInRegOpcode(unsigned Opcode) { diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp index 8615d3060b8e8..6736ebcb513a3 100644 --- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -13008,7 +13008,7 @@ SDValue DAGCombiner::foldPartialReduceMLAMulOp(SDNode *N) { SDValue LHS = Op1->getOperand(0); SDValue RHS = Op1->getOperand(1); unsigned LHSOpcode = LHS->getOpcode(); - if (!ISD::isExtOpcode(LHSOpcode)) + if (!ISD::isExtOpcode(LHSOpcode) && LHSOpcode != ISD::FP_EXTEND) return SDValue(); SDValue LHSExtOp = LHS->getOperand(0); @@ -13040,7 +13040,7 @@ SDValue DAGCombiner::foldPartialReduceMLAMulOp(SDNode *N) { } unsigned RHSOpcode = RHS->getOpcode(); - if (!ISD::isExtOpcode(RHSOpcode)) + if (!ISD::isExtOpcode(RHSOpcode) && RHSOpcode != ISD::FP_EXTEND) return SDValue(); SDValue RHSExtOp = RHS->getOperand(0); From 7683d0e19ca326801f765d514c757618afebe503 Mon Sep 17 00:00:00 2001 From: Damian Heaton Date: Mon, 22 Sep 2025 10:15:46 +0000 Subject: [PATCH 3/7] Address review comments Corrected LangRef typos, improved const comparisons for fadd, and add direct tests. --- llvm/docs/LangRef.rst | 6 +- llvm/include/llvm/IR/Intrinsics.td | 4 +- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 20 +++--- .../CodeGen/SelectionDAG/TargetLowering.cpp | 8 ++- .../lib/Target/AArch64/AArch64SVEInstrInfo.td | 3 - llvm/lib/Target/AArch64/SVEInstrFormats.td | 1 + llvm/test/CodeGen/AArch64/sve2p1-fdot.ll | 66 +++++++++++++++++++ 7 files changed, 90 insertions(+), 18 deletions(-) diff --git a/llvm/docs/LangRef.rst b/llvm/docs/LangRef.rst index d89201477e4b1..c2172c9a2defa 100644 --- a/llvm/docs/LangRef.rst +++ b/llvm/docs/LangRef.rst @@ -20741,8 +20741,8 @@ This is an overloaded intrinsic. :: - declare <4 x f32> @llvm.vector.partial.reduce.fadd.v4f32.v4f32.v8f32(<4 x f32> %a, <8 x f32> %b) - declare @llvm.vector.partial.reduce.add.nxv4f32.nxv4f32.nxv8f32( %a, %b) + declare <4 x f32> @llvm.vector.partial.reduce.fadd.v4f32.v8f32(<4 x f32> %a, <8 x f32> %b) + declare @llvm.vector.partial.reduce.fadd.nxv4f32.nxv8f32( %a, %b) Overview: """"""""" @@ -20762,7 +20762,7 @@ of the result's type, while maintaining the same element type. Semantics: """""""""" -Other than the reduction operator (e.g. add) the way in which the concatenated +Other than the reduction operator (e.g. fadd) the way in which the concatenated arguments is reduced is entirely unspecified. By their nature these intrinsics are not expected to be useful in isolation but instead implement the first phase of an overall reduction operation. diff --git a/llvm/include/llvm/IR/Intrinsics.td b/llvm/include/llvm/IR/Intrinsics.td index 9ef3705250b46..ce79a58053754 100644 --- a/llvm/include/llvm/IR/Intrinsics.td +++ b/llvm/include/llvm/IR/Intrinsics.td @@ -2803,8 +2803,8 @@ def int_vector_partial_reduce_add : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [IntrNoMem]>; def int_vector_partial_reduce_fadd : DefaultAttrsIntrinsic<[LLVMMatchType<0>], - [llvm_anyfloat_ty, llvm_anyfloat_ty], - [IntrNoMem]>; + [llvm_anyfloat_ty, llvm_anyfloat_ty], + [IntrNoMem]>; //===----------------- Pointer Authentication Intrinsics ------------------===// // diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp index 6736ebcb513a3..220220c01272b 100644 --- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -12998,11 +12998,12 @@ SDValue DAGCombiner::foldPartialReduceMLAMulOp(SDNode *N) { SDValue Op2 = N->getOperand(2); APInt C; + ConstantFPSDNode *CFP; if (!(Op1->getOpcode() == ISD::MUL && ISD::isConstantSplatVector(Op2.getNode(), C) && C.isOne()) && !(Op1->getOpcode() == ISD::FMUL && - ISD::isConstantSplatVector(Op2.getNode(), C) && - C == APFloat(1.0f).bitcastToAPInt().trunc(C.getBitWidth()))) + (CFP = llvm::isConstOrConstSplatFP(Op2, false)) && + CFP->isExactlyValue(1.0))) return SDValue(); SDValue LHS = Op1->getOperand(0); @@ -13093,20 +13094,23 @@ SDValue DAGCombiner::foldPartialReduceAdd(SDNode *N) { SDValue Op2 = N->getOperand(2); APInt ConstantOne; - if (!ISD::isConstantSplatVector(Op2.getNode(), ConstantOne) || - !(ConstantOne.isOne() || - ConstantOne == - APFloat(1.0f).bitcastToAPInt().trunc(ConstantOne.getBitWidth()))) + ConstantFPSDNode *C; + if (!(N->getOpcode() == ISD::PARTIAL_REDUCE_FMLA && + (C = llvm::isConstOrConstSplatFP(Op2, false)) && + C->isExactlyValue(1.0)) && + !(ISD::isConstantSplatVector(Op2.getNode(), ConstantOne) && + ConstantOne.isOne())) return SDValue(); unsigned Op1Opcode = Op1.getOpcode(); if (!ISD::isExtOpcode(Op1Opcode)) return SDValue(); - bool Op1IsSigned = Op1Opcode != ISD::ZERO_EXTEND; + bool Op1IsSigned = Op1Opcode == ISD::SIGN_EXTEND; bool NodeIsSigned = N->getOpcode() != ISD::PARTIAL_REDUCE_UMLA; EVT AccElemVT = Acc.getValueType().getVectorElementType(); - if (Op1IsSigned != NodeIsSigned && + if (N->getOpcode() != ISD::PARTIAL_REDUCE_FMLA && + Op1IsSigned != NodeIsSigned && Op1.getValueType().getVectorElementType() != AccElemVT) return SDValue(); diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp index 8e284997af01c..f1b5641db4b5d 100644 --- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp @@ -12077,8 +12077,12 @@ SDValue TargetLowering::expandPartialReduceMLA(SDNode *N, } SDValue Input = MulLHS; APInt ConstantOne; - if (!ISD::isConstantSplatVector(MulRHS.getNode(), ConstantOne) || - !(ConstantOne.isOne() || ConstantOne == APFloat(1.0f).bitcastToAPInt())) + ConstantFPSDNode *C; + if (!(N->getOpcode() == ISD::PARTIAL_REDUCE_FMLA && + (C = llvm::isConstOrConstSplatFP(MulRHS, false)) && + C->isExactlyValue(1.0)) && + !(ISD::isConstantSplatVector(MulRHS.getNode(), ConstantOne) && + ConstantOne.isOne())) Input = DAG.getNode(ISD::MUL, DL, ExtMulOpVT, MulLHS, MulRHS); unsigned Stride = AccVT.getVectorMinNumElements(); diff --git a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td index 9721641d8bde0..1e30735b7a56a 100644 --- a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td +++ b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td @@ -4228,9 +4228,6 @@ defm FCLAMP_ZZZ : sve_fp_clamp<"fclamp", AArch64fclamp>; defm FDOT_ZZZ_S : sve_float_dot<0b0, 0b0, ZPR32, ZPR16, "fdot", nxv8f16, int_aarch64_sve_fdot_x2>; defm FDOT_ZZZI_S : sve_float_dot_indexed<0b0, 0b00, ZPR16, ZPR3b16, "fdot", nxv8f16, int_aarch64_sve_fdot_lane_x2>; -def : Pat<(nxv4f32 (partial_reduce_fmla nxv4f32:$Acc, nxv8f16:$LHS, nxv8f16:$RHS)), - (FDOT_ZZZ_S $Acc, $LHS, $RHS)>; - defm BFMLSLB_ZZZ_S : sve2_fp_mla_long<0b110, "bfmlslb", nxv4f32, nxv8bf16, int_aarch64_sve_bfmlslb>; defm BFMLSLT_ZZZ_S : sve2_fp_mla_long<0b111, "bfmlslt", nxv4f32, nxv8bf16, int_aarch64_sve_bfmlslt>; defm BFMLSLB_ZZZI_S : sve2_fp_mla_long_by_indexed_elem<0b110, "bfmlslb", nxv4f32, nxv8bf16, int_aarch64_sve_bfmlslb_lane>; diff --git a/llvm/lib/Target/AArch64/SVEInstrFormats.td b/llvm/lib/Target/AArch64/SVEInstrFormats.td index 9a23c35766cac..2323ab49dfba3 100644 --- a/llvm/lib/Target/AArch64/SVEInstrFormats.td +++ b/llvm/lib/Target/AArch64/SVEInstrFormats.td @@ -9457,6 +9457,7 @@ multiclass sve_float_dot { def NAME : sve_float_dot; def : SVE_3_Op_Pat(NAME)>; + def : SVE_3_Op_Pat(NAME)>; } multiclass sve_fp8_dot @llvm.vector.partial.reduce.fadd(<4 x float> %acc, <8 x float> %mult) ret <4 x float> %partial.reduce } + +define <8 x half> @partial_reduce_half(<8 x half> %acc, <16 x half> %a) { +; CHECK-LABEL: partial_reduce_half: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: fadd v0.8h, v0.8h, v1.8h +; CHECK-NEXT: fadd v0.8h, v0.8h, v2.8h +; CHECK-NEXT: ret +entry: + %partial.reduce = call <8 x half> @llvm.vector.partial.reduce.fadd(<8 x half> %acc, <16 x half> %a) + ret <8 x half> %partial.reduce +} + +define <4 x float> @partial_reduce_float(<4 x float> %acc, <8 x float> %a) { +; CHECK-LABEL: partial_reduce_float: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: fadd v0.4s, v0.4s, v1.4s +; CHECK-NEXT: fadd v0.4s, v0.4s, v2.4s +; CHECK-NEXT: ret +entry: + %partial.reduce = call <4 x float> @llvm.vector.partial.reduce.fadd(<4 x float> %acc, <8 x float> %a) + ret <4 x float> %partial.reduce +} + +define <2 x double> @partial_reduce_double(<2 x double> %acc, <4 x double> %a) { +; CHECK-LABEL: partial_reduce_double: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: fadd v0.2d, v0.2d, v1.2d +; CHECK-NEXT: fadd v0.2d, v0.2d, v2.2d +; CHECK-NEXT: ret +entry: + %partial.reduce = call <2 x double> @llvm.vector.partial.reduce.fadd(<2 x double> %acc, <4 x double> %a) + ret <2 x double> %partial.reduce +} + +define @partial_reduce_half_vl128( %acc, %a) { +; CHECK-LABEL: partial_reduce_half_vl128: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: fadd z0.h, z0.h, z1.h +; CHECK-NEXT: fadd z0.h, z0.h, z2.h +; CHECK-NEXT: ret +entry: + %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %a) + ret %partial.reduce +} + +define @partial_reduce_float_vl128( %acc, %a) { +; CHECK-LABEL: partial_reduce_float_vl128: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: fadd z0.s, z0.s, z1.s +; CHECK-NEXT: fadd z0.s, z0.s, z2.s +; CHECK-NEXT: ret +entry: + %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %a) + ret %partial.reduce +} + +define @partial_reduce_double_vl128( %acc, %a) { +; CHECK-LABEL: partial_reduce_double_vl128: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: fadd z0.d, z0.d, z1.d +; CHECK-NEXT: fadd z0.d, z0.d, z2.d +; CHECK-NEXT: ret +entry: + %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %a) + ret %partial.reduce +} From 39e962b5f757a92297b6d5b425298cdbc3ef726c Mon Sep 17 00:00:00 2001 From: Damian Heaton Date: Tue, 23 Sep 2025 09:13:39 +0000 Subject: [PATCH 4/7] Require reassoc --- llvm/lib/IR/Verifier.cpp | 6 ++++++ llvm/test/CodeGen/AArch64/sve2p1-fdot.ll | 18 +++++++++--------- 2 files changed, 15 insertions(+), 9 deletions(-) diff --git a/llvm/lib/IR/Verifier.cpp b/llvm/lib/IR/Verifier.cpp index b2e76cc7a8a90..75ddd9421d9eb 100644 --- a/llvm/lib/IR/Verifier.cpp +++ b/llvm/lib/IR/Verifier.cpp @@ -6561,6 +6561,12 @@ void Verifier::visitIntrinsicCall(Intrinsic::ID ID, CallBase &Call) { } break; } + case Intrinsic::vector_partial_reduce_fadd: { + Check(Call.hasAllowReassoc(), + "vector_partial_reduce_fadd requires reassociation to be allowed."); + // Fall through to perform the same verification checks as for integers. + [[fallthrough]]; + } case Intrinsic::vector_partial_reduce_add: { VectorType *AccTy = cast(Call.getArgOperand(0)->getType()); VectorType *VecTy = cast(Call.getArgOperand(1)->getType()); diff --git a/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll b/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll index 69c0b68f23f78..aa2184ab6e65e 100644 --- a/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll +++ b/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll @@ -10,7 +10,7 @@ entry: %a.wide = fpext %a to %b.wide = fpext %b to %mult = fmul %a.wide, %b.wide - %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %mult) + %partial.reduce = call reassoc @llvm.vector.partial.reduce.fadd( %acc, %mult) ret %partial.reduce } @@ -40,7 +40,7 @@ entry: %a.wide = fpext <16 x half> %a to <16 x float> %b.wide = fpext <16 x half> %b to <16 x float> %mult = fmul <16 x float> %a.wide, %b.wide - %partial.reduce = call <8 x float> @llvm.vector.partial.reduce.fadd(<8 x float> %acc, <16 x float> %mult) + %partial.reduce = call reassoc <8 x float> @llvm.vector.partial.reduce.fadd(<8 x float> %acc, <16 x float> %mult) store <8 x float> %partial.reduce, ptr %accptr ret void } @@ -61,7 +61,7 @@ entry: %a.wide = fpext <8 x half> %a to <8 x float> %b.wide = fpext <8 x half> %b to <8 x float> %mult = fmul <8 x float> %a.wide, %b.wide - %partial.reduce = call <4 x float> @llvm.vector.partial.reduce.fadd(<4 x float> %acc, <8 x float> %mult) + %partial.reduce = call reassoc <4 x float> @llvm.vector.partial.reduce.fadd(<4 x float> %acc, <8 x float> %mult) ret <4 x float> %partial.reduce } @@ -72,7 +72,7 @@ define <8 x half> @partial_reduce_half(<8 x half> %acc, <16 x half> %a) { ; CHECK-NEXT: fadd v0.8h, v0.8h, v2.8h ; CHECK-NEXT: ret entry: - %partial.reduce = call <8 x half> @llvm.vector.partial.reduce.fadd(<8 x half> %acc, <16 x half> %a) + %partial.reduce = call reassoc <8 x half> @llvm.vector.partial.reduce.fadd(<8 x half> %acc, <16 x half> %a) ret <8 x half> %partial.reduce } @@ -83,7 +83,7 @@ define <4 x float> @partial_reduce_float(<4 x float> %acc, <8 x float> %a) { ; CHECK-NEXT: fadd v0.4s, v0.4s, v2.4s ; CHECK-NEXT: ret entry: - %partial.reduce = call <4 x float> @llvm.vector.partial.reduce.fadd(<4 x float> %acc, <8 x float> %a) + %partial.reduce = call reassoc <4 x float> @llvm.vector.partial.reduce.fadd(<4 x float> %acc, <8 x float> %a) ret <4 x float> %partial.reduce } @@ -94,7 +94,7 @@ define <2 x double> @partial_reduce_double(<2 x double> %acc, <4 x double> %a) { ; CHECK-NEXT: fadd v0.2d, v0.2d, v2.2d ; CHECK-NEXT: ret entry: - %partial.reduce = call <2 x double> @llvm.vector.partial.reduce.fadd(<2 x double> %acc, <4 x double> %a) + %partial.reduce = call reassoc <2 x double> @llvm.vector.partial.reduce.fadd(<2 x double> %acc, <4 x double> %a) ret <2 x double> %partial.reduce } @@ -105,7 +105,7 @@ define @partial_reduce_half_vl128( %acc, ; CHECK-NEXT: fadd z0.h, z0.h, z2.h ; CHECK-NEXT: ret entry: - %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %a) + %partial.reduce = call reassoc @llvm.vector.partial.reduce.fadd( %acc, %a) ret %partial.reduce } @@ -116,7 +116,7 @@ define @partial_reduce_float_vl128( %ac ; CHECK-NEXT: fadd z0.s, z0.s, z2.s ; CHECK-NEXT: ret entry: - %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %a) + %partial.reduce = call reassoc @llvm.vector.partial.reduce.fadd( %acc, %a) ret %partial.reduce } @@ -127,6 +127,6 @@ define @partial_reduce_double_vl128( ; CHECK-NEXT: fadd z0.d, z0.d, z2.d ; CHECK-NEXT: ret entry: - %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %a) + %partial.reduce = call reassoc @llvm.vector.partial.reduce.fadd( %acc, %a) ret %partial.reduce } From c3bcf04a850e4d62bbd28cd66c69e6cbb2468474 Mon Sep 17 00:00:00 2001 From: Damian Heaton Date: Fri, 26 Sep 2025 09:55:22 +0000 Subject: [PATCH 5/7] Revert "Require reassoc" This reverts commit 319852132602f685aea6228f10418370fd530aa7. --- llvm/lib/IR/Verifier.cpp | 6 ------ llvm/test/CodeGen/AArch64/sve2p1-fdot.ll | 18 +++++++++--------- 2 files changed, 9 insertions(+), 15 deletions(-) diff --git a/llvm/lib/IR/Verifier.cpp b/llvm/lib/IR/Verifier.cpp index 75ddd9421d9eb..b2e76cc7a8a90 100644 --- a/llvm/lib/IR/Verifier.cpp +++ b/llvm/lib/IR/Verifier.cpp @@ -6561,12 +6561,6 @@ void Verifier::visitIntrinsicCall(Intrinsic::ID ID, CallBase &Call) { } break; } - case Intrinsic::vector_partial_reduce_fadd: { - Check(Call.hasAllowReassoc(), - "vector_partial_reduce_fadd requires reassociation to be allowed."); - // Fall through to perform the same verification checks as for integers. - [[fallthrough]]; - } case Intrinsic::vector_partial_reduce_add: { VectorType *AccTy = cast(Call.getArgOperand(0)->getType()); VectorType *VecTy = cast(Call.getArgOperand(1)->getType()); diff --git a/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll b/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll index aa2184ab6e65e..69c0b68f23f78 100644 --- a/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll +++ b/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll @@ -10,7 +10,7 @@ entry: %a.wide = fpext %a to %b.wide = fpext %b to %mult = fmul %a.wide, %b.wide - %partial.reduce = call reassoc @llvm.vector.partial.reduce.fadd( %acc, %mult) + %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %mult) ret %partial.reduce } @@ -40,7 +40,7 @@ entry: %a.wide = fpext <16 x half> %a to <16 x float> %b.wide = fpext <16 x half> %b to <16 x float> %mult = fmul <16 x float> %a.wide, %b.wide - %partial.reduce = call reassoc <8 x float> @llvm.vector.partial.reduce.fadd(<8 x float> %acc, <16 x float> %mult) + %partial.reduce = call <8 x float> @llvm.vector.partial.reduce.fadd(<8 x float> %acc, <16 x float> %mult) store <8 x float> %partial.reduce, ptr %accptr ret void } @@ -61,7 +61,7 @@ entry: %a.wide = fpext <8 x half> %a to <8 x float> %b.wide = fpext <8 x half> %b to <8 x float> %mult = fmul <8 x float> %a.wide, %b.wide - %partial.reduce = call reassoc <4 x float> @llvm.vector.partial.reduce.fadd(<4 x float> %acc, <8 x float> %mult) + %partial.reduce = call <4 x float> @llvm.vector.partial.reduce.fadd(<4 x float> %acc, <8 x float> %mult) ret <4 x float> %partial.reduce } @@ -72,7 +72,7 @@ define <8 x half> @partial_reduce_half(<8 x half> %acc, <16 x half> %a) { ; CHECK-NEXT: fadd v0.8h, v0.8h, v2.8h ; CHECK-NEXT: ret entry: - %partial.reduce = call reassoc <8 x half> @llvm.vector.partial.reduce.fadd(<8 x half> %acc, <16 x half> %a) + %partial.reduce = call <8 x half> @llvm.vector.partial.reduce.fadd(<8 x half> %acc, <16 x half> %a) ret <8 x half> %partial.reduce } @@ -83,7 +83,7 @@ define <4 x float> @partial_reduce_float(<4 x float> %acc, <8 x float> %a) { ; CHECK-NEXT: fadd v0.4s, v0.4s, v2.4s ; CHECK-NEXT: ret entry: - %partial.reduce = call reassoc <4 x float> @llvm.vector.partial.reduce.fadd(<4 x float> %acc, <8 x float> %a) + %partial.reduce = call <4 x float> @llvm.vector.partial.reduce.fadd(<4 x float> %acc, <8 x float> %a) ret <4 x float> %partial.reduce } @@ -94,7 +94,7 @@ define <2 x double> @partial_reduce_double(<2 x double> %acc, <4 x double> %a) { ; CHECK-NEXT: fadd v0.2d, v0.2d, v2.2d ; CHECK-NEXT: ret entry: - %partial.reduce = call reassoc <2 x double> @llvm.vector.partial.reduce.fadd(<2 x double> %acc, <4 x double> %a) + %partial.reduce = call <2 x double> @llvm.vector.partial.reduce.fadd(<2 x double> %acc, <4 x double> %a) ret <2 x double> %partial.reduce } @@ -105,7 +105,7 @@ define @partial_reduce_half_vl128( %acc, ; CHECK-NEXT: fadd z0.h, z0.h, z2.h ; CHECK-NEXT: ret entry: - %partial.reduce = call reassoc @llvm.vector.partial.reduce.fadd( %acc, %a) + %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %a) ret %partial.reduce } @@ -116,7 +116,7 @@ define @partial_reduce_float_vl128( %ac ; CHECK-NEXT: fadd z0.s, z0.s, z2.s ; CHECK-NEXT: ret entry: - %partial.reduce = call reassoc @llvm.vector.partial.reduce.fadd( %acc, %a) + %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %a) ret %partial.reduce } @@ -127,6 +127,6 @@ define @partial_reduce_double_vl128( ; CHECK-NEXT: fadd z0.d, z0.d, z2.d ; CHECK-NEXT: ret entry: - %partial.reduce = call reassoc @llvm.vector.partial.reduce.fadd( %acc, %a) + %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %a) ret %partial.reduce } From 596edec7139fb51fa65f1eb58104a8dadedba9b2 Mon Sep 17 00:00:00 2001 From: Damian Heaton Date: Wed, 8 Oct 2025 10:38:25 +0000 Subject: [PATCH 6/7] Re-add Verifier case, and consolidate partial reduction docs --- llvm/docs/LangRef.rst | 64 +++++++++++++++------------------------- llvm/lib/IR/Verifier.cpp | 1 + 2 files changed, 24 insertions(+), 41 deletions(-) diff --git a/llvm/docs/LangRef.rst b/llvm/docs/LangRef.rst index c2172c9a2defa..99d37a28233ed 100644 --- a/llvm/docs/LangRef.rst +++ b/llvm/docs/LangRef.rst @@ -20688,8 +20688,27 @@ Note that it has the following implications: - If ``%cnt`` is non-zero, the return value is non-zero as well. - If ``%cnt`` is less than or equal to ``%max_lanes``, the return value is equal to ``%cnt``. +Vector Partial Reduction Intrinsics +----------------------------------- + +Partial horizontal reductions of vectors can be expressed using the following intrinsics. +Each one reduces the concatenation of the two vector arguments down to the number of elements +of the result vector type. + +Other than the reduction operator (e.g. add, fadd) the way in which the concatenated +arguments is reduced is entirely unspecified. By their nature these intrinsics +are not expected to be useful in isolation but instead implement the first phase +of an overall reduction operation. + +The typical use case is loop vectorization where reductions are split into an +in-loop phase, where maintaining an unordered vector result is important for +performance, and an out-of-loop phase to calculate the final scalar result. + +By avoiding the introduction of new ordering constraints, these intrinsics +enhance the ability to leverage a target's accumulation instructions. + '``llvm.vector.partial.reduce.add.*``' Intrinsic -^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Syntax: """"""" @@ -20702,13 +20721,6 @@ This is an overloaded intrinsic. declare @llvm.vector.partial.reduce.add.nxv4i32.nxv4i32.nxv8i32( %a, %b) declare @llvm.vector.partial.reduce.add.nxv4i32.nxv4i32.nxv16i32( %a, %b) -Overview: -""""""""" - -The '``llvm.vector.partial.reduce.add.*``' intrinsics reduce the -concatenation of the two vector arguments down to the number of elements of the -result vector type. - Arguments: """""""""" @@ -20717,21 +20729,6 @@ The first argument is an integer vector with the same type as the result. The second argument is a vector with a length that is a known integer multiple of the result's type, while maintaining the same element type. -Semantics: -"""""""""" - -Other than the reduction operator (e.g. add) the way in which the concatenated -arguments is reduced is entirely unspecified. By their nature these intrinsics -are not expected to be useful in isolation but instead implement the first phase -of an overall reduction operation. - -The typical use case is loop vectorization where reductions are split into an -in-loop phase, where maintaining an unordered vector result is important for -performance, and an out-of-loop phase to calculate the final scalar result. - -By avoiding the introduction of new ordering constraints, these intrinsics -enhance the ability to leverage a target's accumulation instructions. - '``llvm.vector.partial.reduce.fadd.*``' Intrinsic ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ @@ -20744,13 +20741,6 @@ This is an overloaded intrinsic. declare <4 x f32> @llvm.vector.partial.reduce.fadd.v4f32.v8f32(<4 x f32> %a, <8 x f32> %b) declare @llvm.vector.partial.reduce.fadd.nxv4f32.nxv8f32( %a, %b) -Overview: -""""""""" - -The '``llvm.vector.partial.reduce.fadd.*``' intrinsics reduce the -concatenation of the two vector arguments down to the number of elements of the -result vector type. - Arguments: """""""""" @@ -20762,17 +20752,9 @@ of the result's type, while maintaining the same element type. Semantics: """""""""" -Other than the reduction operator (e.g. fadd) the way in which the concatenated -arguments is reduced is entirely unspecified. By their nature these intrinsics -are not expected to be useful in isolation but instead implement the first phase -of an overall reduction operation. - -The typical use case is loop vectorization where reductions are split into an -in-loop phase, where maintaining an unordered vector result is important for -performance, and an out-of-loop phase to calculate the final scalar result. - -By avoiding the introduction of new ordering constraints, these intrinsics -enhance the ability to leverage a target's accumulation instructions. +As the way in which the arguments to this floating-point intrinsic are reduced is unspecified, +this intrinsic will reassociate floating-point values, which may result in variations to the +results due to reordering or by lowering to different instructions. '``llvm.experimental.vector.histogram.*``' Intrinsic ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ diff --git a/llvm/lib/IR/Verifier.cpp b/llvm/lib/IR/Verifier.cpp index b2e76cc7a8a90..7c3ca6e4574f3 100644 --- a/llvm/lib/IR/Verifier.cpp +++ b/llvm/lib/IR/Verifier.cpp @@ -6561,6 +6561,7 @@ void Verifier::visitIntrinsicCall(Intrinsic::ID ID, CallBase &Call) { } break; } + case Intrinsic::vector_partial_reduce_fadd: case Intrinsic::vector_partial_reduce_add: { VectorType *AccTy = cast(Call.getArgOperand(0)->getType()); VectorType *VecTy = cast(Call.getArgOperand(1)->getType()); From 41d845fccda0e0a0a9cce62d1077c4ef26419106 Mon Sep 17 00:00:00 2001 From: Damian Heaton Date: Fri, 10 Oct 2025 13:20:42 +0000 Subject: [PATCH 7/7] Address review comments --- llvm/docs/LangRef.rst | 138 +++++++++--------- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 24 ++- llvm/test/CodeGen/AArch64/sve2p1-fdot.ll | 12 ++ 3 files changed, 102 insertions(+), 72 deletions(-) diff --git a/llvm/docs/LangRef.rst b/llvm/docs/LangRef.rst index 99d37a28233ed..2066167345945 100644 --- a/llvm/docs/LangRef.rst +++ b/llvm/docs/LangRef.rst @@ -20315,6 +20315,76 @@ Arguments: """""""""" The argument to this intrinsic must be a vector of floating-point values. +Vector Partial Reduction Intrinsics +----------------------------------- + +Partial reductions of vectors can be expressed using the following intrinsics. +Each one reduces the concatenation of the two vector arguments down to the +number of elements of the result vector type. + +Other than the reduction operator (e.g. add, fadd) the way in which the +concatenated arguments is reduced is entirely unspecified. By their nature these +intrinsics are not expected to be useful in isolation but instead implement the +first phase of an overall reduction operation. + +The typical use case is loop vectorization where reductions are split into an +in-loop phase, where maintaining an unordered vector result is important for +performance, and an out-of-loop phase to calculate the final scalar result. + +By avoiding the introduction of new ordering constraints, these intrinsics +enhance the ability to leverage a target's accumulation instructions. + +'``llvm.vector.partial.reduce.add.*``' Intrinsic +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + +Syntax: +""""""" +This is an overloaded intrinsic. + +:: + + declare <4 x i32> @llvm.vector.partial.reduce.add.v4i32.v4i32.v8i32(<4 x i32> %a, <8 x i32> %b) + declare <4 x i32> @llvm.vector.partial.reduce.add.v4i32.v4i32.v16i32(<4 x i32> %a, <16 x i32> %b) + declare @llvm.vector.partial.reduce.add.nxv4i32.nxv4i32.nxv8i32( %a, %b) + declare @llvm.vector.partial.reduce.add.nxv4i32.nxv4i32.nxv16i32( %a, %b) + +Arguments: +"""""""""" + +The first argument is an integer vector with the same type as the result. + +The second argument is a vector with a length that is a known integer multiple +of the result's type, while maintaining the same element type. + +'``llvm.vector.partial.reduce.fadd.*``' Intrinsic +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + +Syntax: +""""""" +This is an overloaded intrinsic. + +:: + + declare <4 x f32> @llvm.vector.partial.reduce.fadd.v4f32.v8f32(<4 x f32> %a, <8 x f32> %b) + declare @llvm.vector.partial.reduce.fadd.nxv4f32.nxv8f32( %a, %b) + +Arguments: +"""""""""" + +The first argument is a floating-point vector with the same type as the result. + +The second argument is a vector with a length that is a known integer multiple +of the result's type, while maintaining the same element type. + +Semantics: +"""""""""" + +As the way in which the arguments to this floating-point intrinsic are reduced +is unspecified, this intrinsic will assume floating-point reassociation and +contraction, which may result in variations to the results due to reordering or +by lowering to different instructions (including combining multiple instructions +into a single one). + '``llvm.vector.insert``' Intrinsic ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ @@ -20688,74 +20758,6 @@ Note that it has the following implications: - If ``%cnt`` is non-zero, the return value is non-zero as well. - If ``%cnt`` is less than or equal to ``%max_lanes``, the return value is equal to ``%cnt``. -Vector Partial Reduction Intrinsics ------------------------------------ - -Partial horizontal reductions of vectors can be expressed using the following intrinsics. -Each one reduces the concatenation of the two vector arguments down to the number of elements -of the result vector type. - -Other than the reduction operator (e.g. add, fadd) the way in which the concatenated -arguments is reduced is entirely unspecified. By their nature these intrinsics -are not expected to be useful in isolation but instead implement the first phase -of an overall reduction operation. - -The typical use case is loop vectorization where reductions are split into an -in-loop phase, where maintaining an unordered vector result is important for -performance, and an out-of-loop phase to calculate the final scalar result. - -By avoiding the introduction of new ordering constraints, these intrinsics -enhance the ability to leverage a target's accumulation instructions. - -'``llvm.vector.partial.reduce.add.*``' Intrinsic -^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ - -Syntax: -""""""" -This is an overloaded intrinsic. - -:: - - declare <4 x i32> @llvm.vector.partial.reduce.add.v4i32.v4i32.v8i32(<4 x i32> %a, <8 x i32> %b) - declare <4 x i32> @llvm.vector.partial.reduce.add.v4i32.v4i32.v16i32(<4 x i32> %a, <16 x i32> %b) - declare @llvm.vector.partial.reduce.add.nxv4i32.nxv4i32.nxv8i32( %a, %b) - declare @llvm.vector.partial.reduce.add.nxv4i32.nxv4i32.nxv16i32( %a, %b) - -Arguments: -"""""""""" - -The first argument is an integer vector with the same type as the result. - -The second argument is a vector with a length that is a known integer multiple -of the result's type, while maintaining the same element type. - -'``llvm.vector.partial.reduce.fadd.*``' Intrinsic -^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ - -Syntax: -""""""" -This is an overloaded intrinsic. - -:: - - declare <4 x f32> @llvm.vector.partial.reduce.fadd.v4f32.v8f32(<4 x f32> %a, <8 x f32> %b) - declare @llvm.vector.partial.reduce.fadd.nxv4f32.nxv8f32( %a, %b) - -Arguments: -"""""""""" - -The first argument is a floating-point vector with the same type as the result. - -The second argument is a vector with a length that is a known integer multiple -of the result's type, while maintaining the same element type. - -Semantics: -"""""""""" - -As the way in which the arguments to this floating-point intrinsic are reduced is unspecified, -this intrinsic will reassociate floating-point values, which may result in variations to the -results due to reordering or by lowering to different instructions. - '``llvm.experimental.vector.histogram.*``' Intrinsic ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp index 220220c01272b..b3228d904ee1e 100644 --- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -12990,6 +12990,12 @@ SDValue DAGCombiner::visitPARTIAL_REDUCE_MLA(SDNode *N) { // // partial_reduce_*mla(acc, mul(ext(x), splat(C)), splat(1)) // -> partial_reduce_*mla(acc, x, C) +// +// partial_reduce_fmla(acc, fmul(fpext(a), fpext(b)), splat(1.0)) +// -> partial_reduce_fmla(acc, a, b) +// +// partial_reduce_fmla(acc, fmul(fpext(x), splat(C)), splat(1.0)) +// -> partial_reduce_fmla(acc, x, C) SDValue DAGCombiner::foldPartialReduceMLAMulOp(SDNode *N) { SDLoc DL(N); auto *Context = DAG.getContext(); @@ -13006,10 +13012,14 @@ SDValue DAGCombiner::foldPartialReduceMLAMulOp(SDNode *N) { CFP->isExactlyValue(1.0))) return SDValue(); + auto IsIntOrFPExtOpcode = [](unsigned int Opcode) { + return (ISD::isExtOpcode(Opcode) || Opcode == ISD::FP_EXTEND); + }; + SDValue LHS = Op1->getOperand(0); SDValue RHS = Op1->getOperand(1); unsigned LHSOpcode = LHS->getOpcode(); - if (!ISD::isExtOpcode(LHSOpcode) && LHSOpcode != ISD::FP_EXTEND) + if (!IsIntOrFPExtOpcode(LHSOpcode)) return SDValue(); SDValue LHSExtOp = LHS->getOperand(0); @@ -13041,7 +13051,7 @@ SDValue DAGCombiner::foldPartialReduceMLAMulOp(SDNode *N) { } unsigned RHSOpcode = RHS->getOpcode(); - if (!ISD::isExtOpcode(RHSOpcode) && RHSOpcode != ISD::FP_EXTEND) + if (!IsIntOrFPExtOpcode(RHSOpcode)) return SDValue(); SDValue RHSExtOp = RHS->getOperand(0); @@ -13087,6 +13097,8 @@ SDValue DAGCombiner::foldPartialReduceMLAMulOp(SDNode *N) { // -> partial.reduce.smla(acc, op, splat(trunc(1))) // partial.reduce.sumla(acc, sext(op), splat(1)) // -> partial.reduce.smla(acc, op, splat(trunc(1))) +// partial.reduce.fmla(acc, fpext(op), splat(1.0)) +// -> partial.reduce.fmla(acc, op, splat(1.0)) SDValue DAGCombiner::foldPartialReduceAdd(SDNode *N) { SDLoc DL(N); SDValue Acc = N->getOperand(0); @@ -13103,7 +13115,7 @@ SDValue DAGCombiner::foldPartialReduceAdd(SDNode *N) { return SDValue(); unsigned Op1Opcode = Op1.getOpcode(); - if (!ISD::isExtOpcode(Op1Opcode)) + if (!ISD::isExtOpcode(Op1Opcode) && Op1Opcode != ISD::FP_EXTEND) return SDValue(); bool Op1IsSigned = Op1Opcode == ISD::SIGN_EXTEND; @@ -13127,8 +13139,12 @@ SDValue DAGCombiner::foldPartialReduceAdd(SDNode *N) { TLI.getTypeToTransformTo(*Context, UnextOp1VT))) return SDValue(); + auto Constant = N->getOpcode() == ISD::PARTIAL_REDUCE_FMLA + ? DAG.getConstantFP(1, DL, UnextOp1VT) + : DAG.getConstant(1, DL, UnextOp1VT); + return DAG.getNode(NewOpcode, DL, N->getValueType(0), Acc, UnextOp1, - DAG.getConstant(1, DL, UnextOp1VT)); + Constant); } SDValue DAGCombiner::visitVP_STRIDED_LOAD(SDNode *N) { diff --git a/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll b/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll index 69c0b68f23f78..5e55a1869e95f 100644 --- a/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll +++ b/llvm/test/CodeGen/AArch64/sve2p1-fdot.ll @@ -14,6 +14,18 @@ entry: ret %partial.reduce } +define @fdot_splat_vl128( %acc, %a) { +; CHECK-LABEL: fdot_splat_vl128: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: fmov z2.h, #1.00000000 +; CHECK-NEXT: fdot z0.s, z1.h, z2.h +; CHECK-NEXT: ret +entry: + %a.wide = fpext %a to + %partial.reduce = call @llvm.vector.partial.reduce.fadd( %acc, %a.wide) + ret %partial.reduce +} + define void @fdot_wide_vl256(ptr %accptr, ptr %aptr, ptr %bptr) vscale_range(2,2) { ; CHECK-LABEL: fdot_wide_vl256: ; CHECK: // %bb.0: // %entry