1 //===- HexagonVExtract.cpp ------------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // This pass will replace multiple occurrences of V6_extractw from the same 9 // vector register with a combination of a vector store and scalar loads. 10 //===----------------------------------------------------------------------===// 11 12 #include "Hexagon.h" 13 #include "HexagonInstrInfo.h" 14 #include "HexagonMachineFunctionInfo.h" 15 #include "HexagonRegisterInfo.h" 16 #include "HexagonSubtarget.h" 17 #include "llvm/ADT/SmallVector.h" 18 #include "llvm/PassSupport.h" 19 #include "llvm/CodeGen/MachineBasicBlock.h" 20 #include "llvm/CodeGen/MachineFunction.h" 21 #include "llvm/CodeGen/MachineFunctionPass.h" 22 #include "llvm/CodeGen/MachineInstrBuilder.h" 23 #include "llvm/CodeGen/MachineRegisterInfo.h" 24 #include "llvm/Support/CommandLine.h" 25 26 #include <map> 27 28 using namespace llvm; 29 30 static cl::opt<unsigned> VExtractThreshold("hexagon-vextract-threshold", 31 cl::Hidden, cl::ZeroOrMore, cl::init(1), 32 cl::desc("Threshold for triggering vextract replacement")); 33 34 namespace llvm { 35 void initializeHexagonVExtractPass(PassRegistry& Registry); 36 FunctionPass *createHexagonVExtract(); 37 } 38 39 namespace { 40 class HexagonVExtract : public MachineFunctionPass { 41 public: 42 static char ID; 43 HexagonVExtract() : MachineFunctionPass(ID) {} 44 45 StringRef getPassName() const override { 46 return "Hexagon optimize vextract"; 47 } 48 void getAnalysisUsage(AnalysisUsage &AU) const override { 49 MachineFunctionPass::getAnalysisUsage(AU); 50 } 51 bool runOnMachineFunction(MachineFunction &MF) override; 52 53 private: 54 const HexagonSubtarget *HST = nullptr; 55 const HexagonInstrInfo *HII = nullptr; 56 57 unsigned genElemLoad(MachineInstr *ExtI, unsigned BaseR, 58 MachineRegisterInfo &MRI); 59 }; 60 61 char HexagonVExtract::ID = 0; 62 } 63 64 INITIALIZE_PASS(HexagonVExtract, "hexagon-vextract", 65 "Hexagon optimize vextract", false, false) 66 67 unsigned HexagonVExtract::genElemLoad(MachineInstr *ExtI, unsigned BaseR, 68 MachineRegisterInfo &MRI) { 69 MachineBasicBlock &ExtB = *ExtI->getParent(); 70 DebugLoc DL = ExtI->getDebugLoc(); 71 Register ElemR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 72 73 Register ExtIdxR = ExtI->getOperand(2).getReg(); 74 unsigned ExtIdxS = ExtI->getOperand(2).getSubReg(); 75 76 // Simplified check for a compile-time constant value of ExtIdxR. 77 if (ExtIdxS == 0) { 78 MachineInstr *DI = MRI.getVRegDef(ExtIdxR); 79 if (DI->getOpcode() == Hexagon::A2_tfrsi) { 80 unsigned V = DI->getOperand(1).getImm(); 81 V &= (HST->getVectorLength()-1) & -4u; 82 83 BuildMI(ExtB, ExtI, DL, HII->get(Hexagon::L2_loadri_io), ElemR) 84 .addReg(BaseR) 85 .addImm(V); 86 return ElemR; 87 } 88 } 89 90 Register IdxR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 91 BuildMI(ExtB, ExtI, DL, HII->get(Hexagon::A2_andir), IdxR) 92 .add(ExtI->getOperand(2)) 93 .addImm(-4); 94 BuildMI(ExtB, ExtI, DL, HII->get(Hexagon::L4_loadri_rr), ElemR) 95 .addReg(BaseR) 96 .addReg(IdxR) 97 .addImm(0); 98 return ElemR; 99 } 100 101 bool HexagonVExtract::runOnMachineFunction(MachineFunction &MF) { 102 HST = &MF.getSubtarget<HexagonSubtarget>(); 103 HII = HST->getInstrInfo(); 104 const auto &HRI = *HST->getRegisterInfo(); 105 MachineRegisterInfo &MRI = MF.getRegInfo(); 106 MachineFrameInfo &MFI = MF.getFrameInfo(); 107 Register AR = 108 MF.getInfo<HexagonMachineFunctionInfo>()->getStackAlignBaseVReg(); 109 std::map<unsigned, SmallVector<MachineInstr*,4>> VExtractMap; 110 unsigned MaxAlign = 0; 111 bool Changed = false; 112 113 for (MachineBasicBlock &MBB : MF) { 114 for (MachineInstr &MI : MBB) { 115 unsigned Opc = MI.getOpcode(); 116 if (Opc != Hexagon::V6_extractw) 117 continue; 118 Register VecR = MI.getOperand(1).getReg(); 119 VExtractMap[VecR].push_back(&MI); 120 } 121 } 122 123 auto EmitAddr = [&] (MachineBasicBlock &BB, MachineBasicBlock::iterator At, 124 DebugLoc dl, int FI, unsigned Offset) { 125 Register AddrR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 126 unsigned FiOpc = AR != 0 ? Hexagon::PS_fia : Hexagon::PS_fi; 127 auto MIB = BuildMI(BB, At, dl, HII->get(FiOpc), AddrR); 128 if (AR) 129 MIB.addReg(AR); 130 MIB.addFrameIndex(FI).addImm(Offset); 131 return AddrR; 132 }; 133 134 for (auto &P : VExtractMap) { 135 unsigned VecR = P.first; 136 if (P.second.size() <= VExtractThreshold) 137 continue; 138 139 const auto &VecRC = *MRI.getRegClass(VecR); 140 unsigned Align = HRI.getSpillAlignment(VecRC); 141 MaxAlign = std::max(MaxAlign, Align); 142 // Make sure this is not a spill slot: spill slots cannot be aligned 143 // if there are variable-sized objects on the stack. They must be 144 // accessible via FP (which is not aligned), because SP is unknown, 145 // and AP may not be available at the location of the load/store. 146 int FI = MFI.CreateStackObject(HRI.getSpillSize(VecRC), Align, 147 /*isSpillSlot*/false); 148 149 MachineInstr *DefI = MRI.getVRegDef(VecR); 150 MachineBasicBlock::iterator At = std::next(DefI->getIterator()); 151 MachineBasicBlock &DefB = *DefI->getParent(); 152 unsigned StoreOpc = VecRC.getID() == Hexagon::HvxVRRegClassID 153 ? Hexagon::V6_vS32b_ai 154 : Hexagon::PS_vstorerw_ai; 155 Register AddrR = EmitAddr(DefB, At, DefI->getDebugLoc(), FI, 0); 156 BuildMI(DefB, At, DefI->getDebugLoc(), HII->get(StoreOpc)) 157 .addReg(AddrR) 158 .addImm(0) 159 .addReg(VecR); 160 161 unsigned VecSize = HRI.getRegSizeInBits(VecRC) / 8; 162 163 for (MachineInstr *ExtI : P.second) { 164 assert(ExtI->getOpcode() == Hexagon::V6_extractw); 165 unsigned SR = ExtI->getOperand(1).getSubReg(); 166 assert(ExtI->getOperand(1).getReg() == VecR); 167 168 MachineBasicBlock &ExtB = *ExtI->getParent(); 169 DebugLoc DL = ExtI->getDebugLoc(); 170 Register BaseR = EmitAddr(ExtB, ExtI, ExtI->getDebugLoc(), FI, 171 SR == 0 ? 0 : VecSize/2); 172 173 unsigned ElemR = genElemLoad(ExtI, BaseR, MRI); 174 Register ExtR = ExtI->getOperand(0).getReg(); 175 MRI.replaceRegWith(ExtR, ElemR); 176 ExtB.erase(ExtI); 177 Changed = true; 178 } 179 } 180 181 if (AR) { 182 // Update the required stack alignment. 183 MachineInstr *AlignaI = MRI.getVRegDef(AR); 184 assert(AlignaI->getOpcode() == Hexagon::PS_aligna); 185 MachineOperand &Op = AlignaI->getOperand(1); 186 if (MaxAlign > Op.getImm()) 187 Op.setImm(MaxAlign); 188 } 189 190 return Changed; 191 } 192 193 FunctionPass *llvm::createHexagonVExtract() { 194 return new HexagonVExtract(); 195 } 196