1 //===- GCNCreateVOPD.cpp - Create VOPD Instructions ----------------------===// 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 // 9 /// \file 10 /// Combine VALU pairs into VOPD instructions 11 /// Only works on wave32 12 /// Has register requirements, we reject creating VOPD if the requirements are 13 /// not met. 14 /// shouldCombineVOPD mutator in postRA machine scheduler puts candidate 15 /// instructions for VOPD back-to-back 16 /// 17 // 18 //===----------------------------------------------------------------------===// 19 20 #include "AMDGPU.h" 21 #include "GCNSubtarget.h" 22 #include "GCNVOPDUtils.h" 23 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 24 #include "SIInstrInfo.h" 25 #include "Utils/AMDGPUBaseInfo.h" 26 #include "llvm/ADT/SmallVector.h" 27 #include "llvm/ADT/Statistic.h" 28 #include "llvm/CodeGen/MachineBasicBlock.h" 29 #include "llvm/CodeGen/MachineInstr.h" 30 #include "llvm/CodeGen/MachineOperand.h" 31 #include "llvm/Support/Casting.h" 32 #include "llvm/Support/Debug.h" 33 #include <utility> 34 35 #define DEBUG_TYPE "gcn-create-vopd" 36 STATISTIC(NumVOPDCreated, "Number of VOPD Insts Created."); 37 38 using namespace llvm; 39 40 namespace { 41 42 class GCNCreateVOPD : public MachineFunctionPass { 43 private: 44 class VOPDCombineInfo { 45 public: 46 VOPDCombineInfo() = default; VOPDCombineInfo(MachineInstr * First,MachineInstr * Second)47 VOPDCombineInfo(MachineInstr *First, MachineInstr *Second) 48 : FirstMI(First), SecondMI(Second) {} 49 50 MachineInstr *FirstMI; 51 MachineInstr *SecondMI; 52 }; 53 54 public: 55 static char ID; 56 const GCNSubtarget *ST = nullptr; 57 GCNCreateVOPD()58 GCNCreateVOPD() : MachineFunctionPass(ID) {} 59 getAnalysisUsage(AnalysisUsage & AU) const60 void getAnalysisUsage(AnalysisUsage &AU) const override { 61 AU.setPreservesCFG(); 62 MachineFunctionPass::getAnalysisUsage(AU); 63 } 64 getPassName() const65 StringRef getPassName() const override { 66 return "GCN Create VOPD Instructions"; 67 } 68 doReplace(const SIInstrInfo * SII,VOPDCombineInfo & CI)69 bool doReplace(const SIInstrInfo *SII, VOPDCombineInfo &CI) { 70 auto *FirstMI = CI.FirstMI; 71 auto *SecondMI = CI.SecondMI; 72 unsigned Opc1 = FirstMI->getOpcode(); 73 unsigned Opc2 = SecondMI->getOpcode(); 74 unsigned EncodingFamily = 75 AMDGPU::getVOPDEncodingFamily(SII->getSubtarget()); 76 int NewOpcode = 77 AMDGPU::getVOPDFull(AMDGPU::getVOPDOpcode(Opc1), 78 AMDGPU::getVOPDOpcode(Opc2), EncodingFamily); 79 assert(NewOpcode != -1 && 80 "Should have previously determined this as a possible VOPD\n"); 81 82 auto VOPDInst = BuildMI(*FirstMI->getParent(), FirstMI, 83 FirstMI->getDebugLoc(), SII->get(NewOpcode)) 84 .setMIFlags(FirstMI->getFlags() | SecondMI->getFlags()); 85 86 namespace VOPD = AMDGPU::VOPD; 87 MachineInstr *MI[] = {FirstMI, SecondMI}; 88 auto InstInfo = 89 AMDGPU::getVOPDInstInfo(FirstMI->getDesc(), SecondMI->getDesc()); 90 91 for (auto CompIdx : VOPD::COMPONENTS) { 92 auto MCOprIdx = InstInfo[CompIdx].getIndexOfDstInMCOperands(); 93 VOPDInst.add(MI[CompIdx]->getOperand(MCOprIdx)); 94 } 95 96 for (auto CompIdx : VOPD::COMPONENTS) { 97 auto CompSrcOprNum = InstInfo[CompIdx].getCompSrcOperandsNum(); 98 for (unsigned CompSrcIdx = 0; CompSrcIdx < CompSrcOprNum; ++CompSrcIdx) { 99 auto MCOprIdx = InstInfo[CompIdx].getIndexOfSrcInMCOperands(CompSrcIdx); 100 VOPDInst.add(MI[CompIdx]->getOperand(MCOprIdx)); 101 } 102 } 103 104 SII->fixImplicitOperands(*VOPDInst); 105 for (auto CompIdx : VOPD::COMPONENTS) 106 VOPDInst.copyImplicitOps(*MI[CompIdx]); 107 108 LLVM_DEBUG(dbgs() << "VOPD Fused: " << *VOPDInst << " from\tX: " 109 << *CI.FirstMI << "\tY: " << *CI.SecondMI << "\n"); 110 111 for (auto CompIdx : VOPD::COMPONENTS) 112 MI[CompIdx]->eraseFromParent(); 113 114 ++NumVOPDCreated; 115 return true; 116 } 117 runOnMachineFunction(MachineFunction & MF)118 bool runOnMachineFunction(MachineFunction &MF) override { 119 if (skipFunction(MF.getFunction())) 120 return false; 121 ST = &MF.getSubtarget<GCNSubtarget>(); 122 if (!AMDGPU::hasVOPD(*ST) || !ST->isWave32()) 123 return false; 124 LLVM_DEBUG(dbgs() << "CreateVOPD Pass:\n"); 125 126 const SIInstrInfo *SII = ST->getInstrInfo(); 127 bool Changed = false; 128 129 SmallVector<VOPDCombineInfo> ReplaceCandidates; 130 131 for (auto &MBB : MF) { 132 auto MII = MBB.begin(), E = MBB.end(); 133 while (MII != E) { 134 auto *FirstMI = &*MII; 135 MII = next_nodbg(MII, MBB.end()); 136 if (MII == MBB.end()) 137 break; 138 if (FirstMI->isDebugInstr()) 139 continue; 140 auto *SecondMI = &*MII; 141 unsigned Opc = FirstMI->getOpcode(); 142 unsigned Opc2 = SecondMI->getOpcode(); 143 llvm::AMDGPU::CanBeVOPD FirstCanBeVOPD = AMDGPU::getCanBeVOPD(Opc); 144 llvm::AMDGPU::CanBeVOPD SecondCanBeVOPD = AMDGPU::getCanBeVOPD(Opc2); 145 VOPDCombineInfo CI; 146 147 if (FirstCanBeVOPD.X && SecondCanBeVOPD.Y) 148 CI = VOPDCombineInfo(FirstMI, SecondMI); 149 else if (FirstCanBeVOPD.Y && SecondCanBeVOPD.X) 150 CI = VOPDCombineInfo(SecondMI, FirstMI); 151 else 152 continue; 153 // checkVOPDRegConstraints cares about program order, but doReplace 154 // cares about X-Y order in the constituted VOPD 155 if (llvm::checkVOPDRegConstraints(*SII, *FirstMI, *SecondMI)) { 156 ReplaceCandidates.push_back(CI); 157 ++MII; 158 } 159 } 160 } 161 for (auto &CI : ReplaceCandidates) { 162 Changed |= doReplace(SII, CI); 163 } 164 165 return Changed; 166 } 167 }; 168 169 } // namespace 170 171 char GCNCreateVOPD::ID = 0; 172 173 char &llvm::GCNCreateVOPDID = GCNCreateVOPD::ID; 174 175 INITIALIZE_PASS(GCNCreateVOPD, DEBUG_TYPE, "GCN Create VOPD Instructions", 176 false, false) 177