lib/Target/AMDGPU/SIFixupVectorISel.cpp

   1 //===-- SIFixupVectorISel.cpp - Fixup post ISel vector issues -------------===//
   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 /// \file
   8 /// SIFixupVectorISel pass cleans up post ISEL Vector issues.
   9 /// Currently this will convert GLOBAL_{LOAD|STORE}_*
  10 /// and GLOBAL_Atomic_* instructions into their _SADDR variants,
  11 /// feeding the sreg into the saddr field of the new instruction.
  12 /// We currently handle a REG_SEQUENCE feeding the vaddr
  13 /// and decompose it into a base and index.
  14 ///
  15 /// Transform:
  16 /// %17:vgpr_32, %19:sreg_64_xexec = V_ADD_I32_e64 %21:sgpr_32, %22:vgpr_32
  17 /// %18:vgpr_32, %20:sreg_64_xexec = V_ADDC_U32_e64 %25:vgpr_32,
  18 ///                                    %24:vgpr_32, %19:sreg_64_xexec
  19 /// %16:vreg_64 = REG_SEQUENCE %17:vgpr_32, %sub0, %18:vgpr_32, %sub1
  20 /// %11:vreg_64 = COPY %16:vreg_64
  21 /// %10:vgpr_32 = GLOBAL_LOAD_DWORD killed %11:vreg_64, 16, 0, 0
  22 /// Into:
  23 /// %4:sreg_64_xexec = S_LOAD_DWORDX2_IMM %1:sgpr_64, 36, 0
  24 /// %14:vreg_64 = REG_SEQUENCE %6:vgpr_32, %sub0, %15:vgpr_32, %sub1
  25 /// %10:vgpr_32 = GLOBAL_LOAD_DWORD_SADDR %14:vreg_64, %4:sreg_64_xexec,16...
  26 ///
  27 //===----------------------------------------------------------------------===//
  28 //
  29
  30 #include "AMDGPU.h"
  31 #include "AMDGPUSubtarget.h"
  32 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
  33 #include "llvm/ADT/Statistic.h"
  34 #include "llvm/CodeGen/MachineFunctionPass.h"
  35 #include "llvm/CodeGen/MachineInstrBuilder.h"
  36 #include "llvm/CodeGen/MachineRegisterInfo.h"
  37 #include "llvm/IR/Function.h"
  38 #include "llvm/IR/LLVMContext.h"
  39 #include "llvm/Support/Debug.h"
  40 #include "llvm/Target/TargetMachine.h"
  41 #define DEBUG_TYPE "si-fixup-vector-isel"
  42
  43 using namespace llvm;
  44
  45 static cl::opt<bool> EnableGlobalSGPRAddr(
  46   "amdgpu-enable-global-sgpr-addr",
  47   cl::desc("Enable use of SGPR regs for GLOBAL LOAD/STORE instructions"),
  48   cl::init(false));
  49
  50 STATISTIC(NumSGPRGlobalOccurs, "Number of global ld/st opportunities");
  51 STATISTIC(NumSGPRGlobalSaddrs, "Number of global sgpr instructions converted");
  52
  53 namespace {
  54
  55 class SIFixupVectorISel : public MachineFunctionPass {
  56 public:
  57   static char ID;
  58
  59 public:
  60   SIFixupVectorISel() : MachineFunctionPass(ID) {
  61     initializeSIFixupVectorISelPass(*PassRegistry::getPassRegistry());
  62   }
  63
  64   bool runOnMachineFunction(MachineFunction &MF) override;
  65
  66   void getAnalysisUsage(AnalysisUsage &AU) const override {
  67     AU.setPreservesCFG();
  68     MachineFunctionPass::getAnalysisUsage(AU);
  69   }
  70 };
  71
  72 } // End anonymous namespace.
  73
  74 INITIALIZE_PASS(SIFixupVectorISel, DEBUG_TYPE,
  75                 "SI Fixup Vector ISel", false, false)
  76
  77 char SIFixupVectorISel::ID = 0;
  78
  79 char &llvm::SIFixupVectorISelID = SIFixupVectorISel::ID;
  80
  81 FunctionPass *llvm::createSIFixupVectorISelPass() {
  82   return new SIFixupVectorISel();
  83 }
  84
  85 static bool findSRegBaseAndIndex(MachineOperand *Op,
  86                                  unsigned &BaseReg,
  87                                  unsigned &IndexReg,
  88                                  MachineRegisterInfo &MRI,
  89                                  const SIRegisterInfo *TRI) {
  90   SmallVector<MachineOperand *, 8> Worklist;
  91   Worklist.push_back(Op);
  92   while (!Worklist.empty()) {
  93     MachineOperand *WOp = Worklist.pop_back_val();
  94     if (!WOp->isReg() || !Register::isVirtualRegister(WOp->getReg()))
  95       continue;
  96     MachineInstr *DefInst = MRI.getUniqueVRegDef(WOp->getReg());
  97     switch (DefInst->getOpcode()) {
  98     default:
  99       continue;
 100     case AMDGPU::COPY:
 101       Worklist.push_back(&DefInst->getOperand(1));
 102       break;
 103     case AMDGPU::REG_SEQUENCE:
 104       if (DefInst->getNumOperands() != 5)
 105         continue;
 106       Worklist.push_back(&DefInst->getOperand(1));
 107       Worklist.push_back(&DefInst->getOperand(3));
 108       break;
 109     case AMDGPU::V_ADD_I32_e64:
 110       // The V_ADD_* and its analogous V_ADDCV_* are generated by
 111       // a previous pass which lowered from an ADD_64_PSEUDO,
 112       // which generates subregs to break up the 64 bit args.
 113       if (DefInst->getOperand(2).getSubReg() != AMDGPU::NoSubRegister)
 114         continue;
 115       BaseReg = DefInst->getOperand(2).getReg();
 116       if (DefInst->getOperand(3).getSubReg() != AMDGPU::NoSubRegister)
 117         continue;
 118       IndexReg = DefInst->getOperand(3).getReg();
 119       // Chase the IndexReg.
 120       MachineInstr *MI = MRI.getUniqueVRegDef(IndexReg);
 121       if (!MI || !MI->isCopy())
 122         continue;
 123       // Make sure the reg class is 64 bit for Index.
 124       // If the Index register is a subreg, we want it to reference
 125       // a 64 bit register which we will use as the Index reg.
 126       const TargetRegisterClass *IdxRC, *BaseRC;
 127       IdxRC = MRI.getRegClass(MI->getOperand(1).getReg());
 128       if (AMDGPU::getRegBitWidth(IdxRC->getID()) != 64)
 129         continue;
 130       IndexReg = MI->getOperand(1).getReg();
 131       // Chase the BaseReg.
 132       MI = MRI.getUniqueVRegDef(BaseReg);
 133       if (!MI || !MI->isCopy())
 134         continue;
 135       // Make sure the register class is 64 bit for Base.
 136       BaseReg = MI->getOperand(1).getReg();
 137       BaseRC = MRI.getRegClass(BaseReg);
 138       if (AMDGPU::getRegBitWidth(BaseRC->getID()) != 64)
 139         continue;
 140       // Make sure Base is SReg and Index is VReg.
 141       if (!TRI->isSGPRReg(MRI, BaseReg))
 142         return false;
 143       if (!TRI->hasVGPRs(MRI.getRegClass(IndexReg)))
 144         return false;
 145       // clear any killed flags on Index and Base regs, used later.
 146       MRI.clearKillFlags(IndexReg);
 147       MRI.clearKillFlags(BaseReg);
 148       return true;
 149     }
 150   }
 151   return false;
 152 }
 153
 154 // Identify Global LOAD|STORE/ATOMIC and try to convert to _SADDR.
 155 static bool fixupGlobalSaddr(MachineBasicBlock &MBB,
 156                              MachineFunction &MF,
 157                              MachineRegisterInfo &MRI,
 158                              const GCNSubtarget &ST,
 159                              const SIInstrInfo *TII,
 160                              const SIRegisterInfo *TRI) {
 161   if (!EnableGlobalSGPRAddr)
 162     return false;
 163   bool FuncModified = false;
 164   MachineBasicBlock::iterator I, Next;
 165   for (I = MBB.begin(); I != MBB.end(); I = Next) {
 166     Next = std::next(I);
 167     MachineInstr &MI = *I;
 168     int NewOpcd = AMDGPU::getGlobalSaddrOp(MI.getOpcode());
 169     if (NewOpcd < 0)
 170       continue;
 171     // Update our statistics on opportunities seen.
 172     ++NumSGPRGlobalOccurs;
 173     LLVM_DEBUG(dbgs() << "Global Mem opp " << MI << '\n');
 174     // Need a Base and Index or we cant transform to _SADDR.
 175     unsigned BaseReg = 0;
 176     unsigned IndexReg = 0;
 177     MachineOperand *Op = TII->getNamedOperand(MI, AMDGPU::OpName::vaddr);
 178     if (!findSRegBaseAndIndex(Op, BaseReg, IndexReg, MRI, TRI))
 179       continue;
 180     ++NumSGPRGlobalSaddrs;
 181     FuncModified = true;
 182     // Create the new _SADDR Memory instruction.
 183     bool HasVdst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst) != nullptr;
 184     MachineOperand *VData = TII->getNamedOperand(MI, AMDGPU::OpName::vdata);
 185     MachineInstr *NewGlob = nullptr;
 186     NewGlob = BuildMI(MBB, I, MI.getDebugLoc(), TII->get(NewOpcd));
 187     if (HasVdst)
 188       NewGlob->addOperand(MF, MI.getOperand(0));
 189     NewGlob->addOperand(MF, MachineOperand::CreateReg(IndexReg, false));
 190     if (VData)
 191       NewGlob->addOperand(MF, *VData);
 192     NewGlob->addOperand(MF, MachineOperand::CreateReg(BaseReg, false));
 193     NewGlob->addOperand(*TII->getNamedOperand(MI, AMDGPU::OpName::offset));
 194
 195     MachineOperand *Glc = TII->getNamedOperand(MI, AMDGPU::OpName::glc);
 196     // Atomics dont have a GLC, so omit the field if not there.
 197     if (Glc)
 198       NewGlob->addOperand(MF, *Glc);
 199
 200     MachineOperand *DLC = TII->getNamedOperand(MI, AMDGPU::OpName::dlc);
 201     if (DLC)
 202       NewGlob->addOperand(MF, *DLC);
 203
 204     NewGlob->addOperand(*TII->getNamedOperand(MI, AMDGPU::OpName::slc));
 205     // _D16 have an vdst_in operand, copy it in.
 206     MachineOperand *VDstInOp = TII->getNamedOperand(MI,
 207                                       AMDGPU::OpName::vdst_in);
 208     if (VDstInOp)
 209       NewGlob->addOperand(MF, *VDstInOp);
 210     NewGlob->copyImplicitOps(MF, MI);
 211     NewGlob->cloneMemRefs(MF, MI);
 212     // Remove the old Global Memop instruction.
 213     MI.eraseFromParent();
 214     LLVM_DEBUG(dbgs() << "New Global Mem " << *NewGlob << '\n');
 215   }
 216   return FuncModified;
 217 }
 218
 219 bool SIFixupVectorISel::runOnMachineFunction(MachineFunction &MF) {
 220   if (skipFunction(MF.getFunction()))
 221     return false;
 222
 223   MachineRegisterInfo &MRI = MF.getRegInfo();
 224   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
 225   const SIInstrInfo *TII = ST.getInstrInfo();
 226   const SIRegisterInfo *TRI = ST.getRegisterInfo();
 227
 228   bool FuncModified = false;
 229   for (MachineBasicBlock &MBB : MF) {
 230     // Cleanup missed Saddr opportunites from ISel.
 231     FuncModified |= fixupGlobalSaddr(MBB, MF, MRI, ST, TII, TRI);
 232   }
 233   return FuncModified;
 234 }