llvm/lib/Target/AMDGPU/SIOptimizeVGPRLiveRange.cpp

//===--------------------- SIOptimizeVGPRLiveRange.cpp  -------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
/// \file
/// This pass tries to remove unnecessary VGPR live ranges in divergent if-else
/// structures and waterfall loops.
///
/// When we do structurization, we usually transform an if-else into two
/// successive if-then (with a flow block to do predicate inversion). Consider a
/// simple case after structurization: A divergent value %a was defined before
/// if-else and used in both THEN (use in THEN is optional) and ELSE part:
///    bb.if:
///      %a = ...
///      ...
///    bb.then:
///      ... = op %a
///      ... // %a can be dead here
///    bb.flow:
///      ...
///    bb.else:
///      ... = %a
///      ...
///    bb.endif
///
///  As register allocator has no idea of the thread-control-flow, it will just
///  assume %a would be alive in the whole range of bb.then because of a later
///  use in bb.else. On AMDGPU architecture, the VGPR is accessed with respect
///  to exec mask. For this if-else case, the lanes active in bb.then will be
///  inactive in bb.else, and vice-versa. So we are safe to say that %a was dead
///  after the last use in bb.then until the end of the block. The reason is
///  the instructions in bb.then will only overwrite lanes that will never be
///  accessed in bb.else.
///
///  This pass aims to tell register allocator that %a is in-fact dead,
///  through inserting a phi-node in bb.flow saying that %a is undef when coming
///  from bb.then, and then replace the uses in the bb.else with the result of
///  newly inserted phi.
///
///  Two key conditions must be met to ensure correctness:
///  1.) The def-point should be in the same loop-level as if-else-endif to make
///      sure the second loop iteration still get correct data.
///  2.) There should be no further uses after the IF-ELSE region.
///
///
/// Waterfall loops get inserted around instructions that use divergent values
/// but can only be executed with a uniform value. For example an indirect call
/// to a divergent address:
///    bb.start:
///      %a = ...
///      %fun = ...
///      ...
///    bb.loop:
///      call %fun (%a)
///      ... // %a can be dead here
///      loop %bb.loop
///
///  The loop block is executed multiple times, but it is run exactly once for
///  each active lane. Similar to the if-else case, the register allocator
///  assumes that %a is live throughout the loop as it is used again in the next
///  iteration. If %a is a VGPR that is unused after the loop, it does not need
///  to be live after its last use in the loop block. By inserting a phi-node at
///  the start of bb.loop that is undef when coming from bb.loop, the register
///  allocation knows that the value of %a does not need to be preserved through
///  iterations of the loop.
///
//
//===----------------------------------------------------------------------===//

#include "SIOptimizeVGPRLiveRange.h"
#include "AMDGPU.h"
#include "GCNSubtarget.h"
#include "SIMachineFunctionInfo.h"
#include "llvm/CodeGen/LiveIntervals.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/IR/Dominators.h"
#include "llvm/InitializePasses.h"

using namespace llvm;

#define DEBUG_TYPE "si-opt-vgpr-liverange"

namespace {

class SIOptimizeVGPRLiveRange {
private:
  const SIRegisterInfo *TRI = nullptr;
  const SIInstrInfo *TII = nullptr;
  LiveIntervals *LIS = nullptr;
  MachineDominatorTree *MDT = nullptr;
  const MachineLoopInfo *Loops = nullptr;
  MachineRegisterInfo *MRI = nullptr;

  // Is \p Reg alive completely through \p MBB (live-in and live-out with no
  // intervening def/kill)?
  bool isLiveThrough(Register Reg, const MachineBasicBlock *MBB) const;

  // Is \p Reg live into \p MBB? This is true when it is live through MBB or
  // killed in MBB. A register only used by PHIs in MBB is not considered live
  // in.
  bool isLiveIntoMBB(Register Reg, const MachineBasicBlock *MBB) const;

public:
  SIOptimizeVGPRLiveRange(LiveIntervals *LIS, MachineDominatorTree *MDT,
                          MachineLoopInfo *Loops)
      : LIS(LIS), MDT(MDT), Loops(Loops) {}
  bool run(MachineFunction &MF);

  MachineBasicBlock *getElseTarget(MachineBasicBlock *MBB) const;

  void collectElseRegionBlocks(MachineBasicBlock *Flow,
                               MachineBasicBlock *Endif,
                               SmallSetVector<MachineBasicBlock *, 16> &) const;

  void
  collectCandidateRegisters(MachineBasicBlock *If, MachineBasicBlock *Flow,
                            MachineBasicBlock *Endif,
                            SmallSetVector<MachineBasicBlock *, 16> &ElseBlocks,
                            SmallVectorImpl<Register> &CandidateRegs) const;

  void collectWaterfallCandidateRegisters(
      MachineBasicBlock *LoopHeader, MachineBasicBlock *LoopEnd,
      SmallSetVector<Register, 16> &CandidateRegs,
      SmallSetVector<MachineBasicBlock *, 2> &Blocks,
      SmallVectorImpl<MachineInstr *> &Instructions) const;

  void
  optimizeLiveRange(Register Reg, MachineBasicBlock *If,
                    MachineBasicBlock *Flow, MachineBasicBlock *Endif,
                    SmallSetVector<MachineBasicBlock *, 16> &ElseBlocks) const;

  void optimizeWaterfallLiveRange(
      Register Reg, MachineBasicBlock *LoopHeader,
      SmallSetVector<MachineBasicBlock *, 2> &LoopBlocks,
      SmallVectorImpl<MachineInstr *> &Instructions) const;
};

class SIOptimizeVGPRLiveRangeLegacy : public MachineFunctionPass {
public:
  static char ID;

  SIOptimizeVGPRLiveRangeLegacy() : MachineFunctionPass(ID) {}

  bool runOnMachineFunction(MachineFunction &MF) override;

  StringRef getPassName() const override {
    return "SI Optimize VGPR LiveRange";
  }

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.setPreservesCFG();
    AU.addRequired<LiveIntervalsWrapperPass>();
    AU.addPreserved<SlotIndexesWrapperPass>();
    AU.addPreserved<LiveIntervalsWrapperPass>();
    AU.addRequired<MachineDominatorTreeWrapperPass>();
    AU.addRequired<MachineLoopInfoWrapperPass>();
    MachineFunctionPass::getAnalysisUsage(AU);
  }

  MachineFunctionProperties getRequiredProperties() const override {
    return MachineFunctionProperties().setIsSSA();
  }

  MachineFunctionProperties getClearedProperties() const override {
    return MachineFunctionProperties().setNoPHIs();
  }
};

} // end anonymous namespace

// Check whether the MBB is a else flow block and get the branching target which
// is the Endif block
MachineBasicBlock *
SIOptimizeVGPRLiveRange::getElseTarget(MachineBasicBlock *MBB) const {
  for (auto &BR : MBB->terminators()) {
    if (BR.getOpcode() == AMDGPU::SI_ELSE)
      return BR.getOperand(2).getMBB();
  }
  return nullptr;
}

bool SIOptimizeVGPRLiveRange::isLiveThrough(
    Register Reg, const MachineBasicBlock *MBB) const {
  const LiveInterval &LI = LIS->getInterval(Reg);
  return LIS->isLiveInToMBB(LI, MBB) && LIS->isLiveOutOfMBB(LI, MBB);
}

bool SIOptimizeVGPRLiveRange::isLiveIntoMBB(
    Register Reg, const MachineBasicBlock *MBB) const {
  const LiveInterval &LI = LIS->getInterval(Reg);
  return LIS->isLiveInToMBB(LI, MBB);
}

void SIOptimizeVGPRLiveRange::collectElseRegionBlocks(
    MachineBasicBlock *Flow, MachineBasicBlock *Endif,
    SmallSetVector<MachineBasicBlock *, 16> &Blocks) const {
  assert(Flow != Endif);

  MachineBasicBlock *MBB = Endif;
  unsigned Cur = 0;
  while (MBB) {
    for (auto *Pred : MBB->predecessors()) {
      if (Pred != Flow)
        Blocks.insert(Pred);
    }

    if (Cur < Blocks.size())
      MBB = Blocks[Cur++];
    else
      MBB = nullptr;
  }

  LLVM_DEBUG({
    dbgs() << "Found Else blocks: ";
    for (auto *MBB : Blocks)
      dbgs() << printMBBReference(*MBB) << ' ';
    dbgs() << '\n';
  });
}

/// Collect the killed registers in the ELSE region which are not alive through
/// the whole THEN region.
void SIOptimizeVGPRLiveRange::collectCandidateRegisters(
    MachineBasicBlock *If, MachineBasicBlock *Flow, MachineBasicBlock *Endif,
    SmallSetVector<MachineBasicBlock *, 16> &ElseBlocks,
    SmallVectorImpl<Register> &CandidateRegs) const {

  SmallSet<Register, 8> KillsInElse;

  for (auto *Else : ElseBlocks) {
    for (auto &MI : Else->instrs()) {
      if (MI.isDebugInstr())
        continue;

      for (auto &MO : MI.operands()) {
        if (!MO.isReg() || !MO.getReg() || MO.isDef())
          continue;

        Register MOReg = MO.getReg();
        // We can only optimize VGPR/AGPR/AV virtual registers.
        if (!MOReg.isVirtual() ||
            !TRI->hasVectorRegisters(MRI->getRegClass(MOReg)))
          continue;

        if (MO.readsReg()) {
          const MachineBasicBlock *DefMBB = MRI->getDefBlock(MOReg);
          // Make sure two conditions are met:
          // a.) the value is defined before/in the IF block
          // b.) should be defined in the same loop-level.
          if ((isLiveThrough(MOReg, If) || DefMBB == If) &&
              Loops->getLoopFor(DefMBB) == Loops->getLoopFor(If)) {
            // Check if the register is live into the endif block. If not,
            // consider it killed in the else region.
            if (!isLiveIntoMBB(MOReg, Endif)) {
              KillsInElse.insert(MOReg);
            } else {
              LLVM_DEBUG(dbgs() << "Excluding " << printReg(MOReg, TRI)
                                << " as Live in Endif\n");
            }
          }
        }
      }
    }
  }

  // Check the phis in the Endif, looking for value coming from the ELSE
  // region. Make sure the phi-use is the last use.
  for (auto &MI : Endif->phis()) {
    for (unsigned Idx = 1; Idx < MI.getNumOperands(); Idx += 2) {
      auto &MO = MI.getOperand(Idx);
      auto *Pred = MI.getOperand(Idx + 1).getMBB();
      if (Pred == Flow)
        continue;
      assert(ElseBlocks.contains(Pred) && "Should be from Else region\n");

      if (!MO.isReg() || !MO.getReg() || MO.isUndef())
        continue;

      Register Reg = MO.getReg();
      if (!Reg.isVirtual() || !TRI->hasVectorRegisters(MRI->getRegClass(Reg)))
        continue;

      if (isLiveIntoMBB(Reg, Endif)) {
        LLVM_DEBUG(dbgs() << "Excluding " << printReg(Reg, TRI)
                          << " as Live in Endif\n");
        continue;
      }
      // Make sure two conditions are met:
      // a.) the value is defined before/in the IF block
      // b.) should be defined in the same loop-level.
      const MachineBasicBlock *DefMBB = MRI->getDefBlock(Reg);
      if ((isLiveThrough(Reg, If) || DefMBB == If) &&
          Loops->getLoopFor(DefMBB) == Loops->getLoopFor(If))
        KillsInElse.insert(Reg);
    }
  }

  auto IsLiveThroughThen = [&](Register Reg) {
    for (auto I = MRI->use_nodbg_begin(Reg), E = MRI->use_nodbg_end(); I != E;
         ++I) {
      if (!I->readsReg())
        continue;
      auto *UseMI = I->getParent();
      auto *UseMBB = UseMI->getParent();
      if (UseMBB == Flow || UseMBB == Endif) {
        if (!UseMI->isPHI())
          return true;

        auto *IncomingMBB = UseMI->getOperand(I.getOperandNo() + 1).getMBB();
        // The register is live through the path If->Flow or Flow->Endif.
        // we should not optimize for such cases.
        if ((UseMBB == Flow && IncomingMBB != If) ||
            (UseMBB == Endif && IncomingMBB == Flow))
          return true;
      }
    }
    return false;
  };

  for (auto Reg : KillsInElse) {
    if (!IsLiveThroughThen(Reg))
      CandidateRegs.push_back(Reg);
  }
}

/// Collect the registers used in the waterfall loop block that are defined
/// before.
void SIOptimizeVGPRLiveRange::collectWaterfallCandidateRegisters(
    MachineBasicBlock *LoopHeader, MachineBasicBlock *LoopEnd,
    SmallSetVector<Register, 16> &CandidateRegs,
    SmallSetVector<MachineBasicBlock *, 2> &Blocks,
    SmallVectorImpl<MachineInstr *> &Instructions) const {

  // Collect loop instructions, potentially spanning multiple blocks
  auto *MBB = LoopHeader;
  for (;;) {
    Blocks.insert(MBB);
    for (auto &MI : *MBB) {
      if (MI.isDebugInstr())
        continue;
      Instructions.push_back(&MI);
    }
    if (MBB == LoopEnd)
      break;

    if ((MBB != LoopHeader && MBB->pred_size() != 1) ||
        (MBB == LoopHeader && MBB->pred_size() != 2) || MBB->succ_size() != 1) {
      LLVM_DEBUG(dbgs() << "Unexpected edges in CFG, ignoring loop\n");
      return;
    }

    MBB = *MBB->succ_begin();
  }

  for (auto *I : Instructions) {
    auto &MI = *I;

    for (auto &MO : MI.all_uses()) {
      if (!MO.getReg())
        continue;

      Register MOReg = MO.getReg();
      // We can only optimize VGPR/AGPR/AV virtual registers.
      if (!MOReg.isVirtual() ||
          !TRI->hasVectorRegisters(MRI->getRegClass(MOReg)))
        continue;

      if (MO.readsReg()) {
        MachineBasicBlock *DefMBB = MRI->getDefBlock(MOReg);
        // Make sure the value is defined before the LOOP block
        if (!Blocks.contains(DefMBB) && !CandidateRegs.contains(MOReg)) {
          // If the variable is used after the loop, the register coalescer will
          // merge the newly created register and remove the phi node again.
          // Just do nothing in that case.
          bool IsUsed = false;
          for (auto *Succ : LoopEnd->successors()) {
            if (!Blocks.contains(Succ) && isLiveIntoMBB(MOReg, Succ)) {
              IsUsed = true;
              break;
            }
          }
          if (!IsUsed) {
            LLVM_DEBUG(dbgs() << "Found candidate reg: "
                              << printReg(MOReg, TRI, 0, MRI) << '\n');
            CandidateRegs.insert(MOReg);
          } else {
            LLVM_DEBUG(dbgs() << "Reg is used after loop, ignoring: "
                              << printReg(MOReg, TRI, 0, MRI) << '\n');
          }
        }
      }
    }
  }
}

void SIOptimizeVGPRLiveRange::optimizeLiveRange(
    Register Reg, MachineBasicBlock *If, MachineBasicBlock *Flow,
    MachineBasicBlock *Endif,
    SmallSetVector<MachineBasicBlock *, 16> &ElseBlocks) const {
  // Insert a new PHI, marking the value from the THEN region being
  // undef.
  LLVM_DEBUG(dbgs() << "Optimizing " << printReg(Reg, TRI) << '\n');
  const auto *RC = MRI->getRegClass(Reg);
  Register NewReg = MRI->createVirtualRegister(RC);
  Register UndefReg = MRI->createVirtualRegister(RC);
  MachineInstrBuilder PHI = BuildMI(*Flow, Flow->getFirstNonPHI(), DebugLoc(),
                                    TII->get(TargetOpcode::PHI), NewReg);
  for (auto *Pred : Flow->predecessors()) {
    if (Pred == If)
      PHI.addReg(Reg).addMBB(Pred);
    else
      PHI.addReg(UndefReg, RegState::Undef).addMBB(Pred);
  }

  // Replace all uses in the ELSE region or the PHIs in ENDIF block
  // Use early increment range because setReg() will update the linked list.
  for (auto &O : make_early_inc_range(MRI->use_operands(Reg))) {
    auto *UseMI = O.getParent();
    auto *UseBlock = UseMI->getParent();
    // Replace uses in Endif block
    if (UseBlock == Endif) {
      if (UseMI->isPHI())
        O.setReg(NewReg);
      else if (UseMI->isDebugInstr())
        continue;
      else {
        // DetectDeadLanes may mark register uses as undef without removing
        // them, in which case a non-phi instruction using the original register
        // may exist in the Endif block even though the register is not live
        // into it.
        assert(!O.readsReg());
      }
      continue;
    }

    // Replace uses in Else region
    if (ElseBlocks.contains(UseBlock))
      O.setReg(NewReg);
  }

  // The new PHI is a def of NewReg and a use of Reg and UndefReg; the uses of
  // Reg in the Else/Endif region were rewritten to NewReg. Kill flags moved
  // with the rewritten operands may no longer mark the last use, so drop them
  // and let the recomputed intervals be the source of truth.
  MRI->clearKillFlags(Reg);
  MRI->clearKillFlags(NewReg);
  LIS->InsertMachineInstrInMaps(*PHI);
  LIS->removeInterval(Reg);
  LIS->createAndComputeVirtRegInterval(Reg);
  LIS->createAndComputeVirtRegInterval(NewReg);
  LIS->createAndComputeVirtRegInterval(UndefReg);
}

void SIOptimizeVGPRLiveRange::optimizeWaterfallLiveRange(
    Register Reg, MachineBasicBlock *LoopHeader,
    SmallSetVector<MachineBasicBlock *, 2> &Blocks,
    SmallVectorImpl<MachineInstr *> &Instructions) const {
  // Insert a new PHI, marking the value from the last loop iteration undef.
  LLVM_DEBUG(dbgs() << "Optimizing " << printReg(Reg, TRI) << '\n');
  const auto *RC = MRI->getRegClass(Reg);
  Register NewReg = MRI->createVirtualRegister(RC);
  Register UndefReg = MRI->createVirtualRegister(RC);

  // Replace all uses in the LOOP region
  // Use early increment range because setReg() will update the linked list.
  for (auto &O : make_early_inc_range(MRI->use_operands(Reg))) {
    auto *UseMI = O.getParent();
    auto *UseBlock = UseMI->getParent();
    // Replace uses in Loop blocks
    if (Blocks.contains(UseBlock))
      O.setReg(NewReg);
  }

  MachineInstrBuilder PHI =
      BuildMI(*LoopHeader, LoopHeader->getFirstNonPHI(), DebugLoc(),
              TII->get(TargetOpcode::PHI), NewReg);
  for (auto *Pred : LoopHeader->predecessors()) {
    if (Blocks.contains(Pred))
      PHI.addReg(UndefReg, RegState::Undef).addMBB(Pred);
    else
      PHI.addReg(Reg).addMBB(Pred);
  }

  LIS->InsertMachineInstrInMaps(*PHI);
  LIS->removeInterval(Reg);
  LIS->createAndComputeVirtRegInterval(Reg);
  LIS->createAndComputeVirtRegInterval(NewReg);
  LIS->createAndComputeVirtRegInterval(UndefReg);
}

char SIOptimizeVGPRLiveRangeLegacy::ID = 0;

INITIALIZE_PASS_BEGIN(SIOptimizeVGPRLiveRangeLegacy, DEBUG_TYPE,
                      "SI Optimize VGPR LiveRange", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)
INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(LiveIntervalsWrapperPass)
INITIALIZE_PASS_END(SIOptimizeVGPRLiveRangeLegacy, DEBUG_TYPE,
                    "SI Optimize VGPR LiveRange", false, false)

char &llvm::SIOptimizeVGPRLiveRangeLegacyID = SIOptimizeVGPRLiveRangeLegacy::ID;

bool SIOptimizeVGPRLiveRangeLegacy::runOnMachineFunction(MachineFunction &MF) {
  if (skipFunction(MF.getFunction()))
    return false;

  LiveIntervals *LIS = &getAnalysis<LiveIntervalsWrapperPass>().getLIS();
  MachineDominatorTree *MDT =
      &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree();
  MachineLoopInfo *Loops = &getAnalysis<MachineLoopInfoWrapperPass>().getLI();
  return SIOptimizeVGPRLiveRange(LIS, MDT, Loops).run(MF);
}

PreservedAnalyses
SIOptimizeVGPRLiveRangePass::run(MachineFunction &MF,
                                 MachineFunctionAnalysisManager &MFAM) {
  MFPropsModifier _(*this, MF);
  LiveIntervals *LIS = &MFAM.getResult<LiveIntervalsAnalysis>(MF);
  MachineDominatorTree *MDT = &MFAM.getResult<MachineDominatorTreeAnalysis>(MF);
  MachineLoopInfo *Loops = &MFAM.getResult<MachineLoopAnalysis>(MF);

  bool Changed = SIOptimizeVGPRLiveRange(LIS, MDT, Loops).run(MF);
  if (!Changed)
    return PreservedAnalyses::all();

  auto PA = getMachineFunctionPassPreservedAnalyses();
  PA.preserve<SlotIndexesAnalysis>();
  PA.preserve<LiveIntervalsAnalysis>();
  PA.preserveSet<CFGAnalyses>();
  return PA;
}

bool SIOptimizeVGPRLiveRange::run(MachineFunction &MF) {
  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
  TII = ST.getInstrInfo();
  TRI = &TII->getRegisterInfo();
  MRI = &MF.getRegInfo();

  bool MadeChange = false;

  // TODO: we need to think about the order of visiting the blocks to get
  // optimal result for nesting if-else cases.
  for (MachineBasicBlock &MBB : MF) {
    for (auto &MI : MBB.terminators()) {
      // Detect the if-else blocks
      if (MI.getOpcode() == AMDGPU::SI_IF) {
        MachineBasicBlock *IfTarget = MI.getOperand(2).getMBB();
        auto *Endif = getElseTarget(IfTarget);
        if (!Endif)
          continue;

        // Skip unexpected control flow.
        if (!MDT->dominates(&MBB, IfTarget) || !MDT->dominates(IfTarget, Endif))
          continue;

        SmallSetVector<MachineBasicBlock *, 16> ElseBlocks;
        SmallVector<Register> CandidateRegs;

        LLVM_DEBUG(dbgs() << "Checking IF-ELSE-ENDIF: "
                          << printMBBReference(MBB) << ' '
                          << printMBBReference(*IfTarget) << ' '
                          << printMBBReference(*Endif) << '\n');

        // Collect all the blocks in the ELSE region
        collectElseRegionBlocks(IfTarget, Endif, ElseBlocks);

        // Collect the registers can be optimized
        collectCandidateRegisters(&MBB, IfTarget, Endif, ElseBlocks,
                                  CandidateRegs);
        MadeChange |= !CandidateRegs.empty();
        // Now we are safe to optimize.
        for (auto Reg : CandidateRegs)
          optimizeLiveRange(Reg, &MBB, IfTarget, Endif, ElseBlocks);
      } else if (MI.getOpcode() == AMDGPU::SI_WATERFALL_LOOP) {
        auto *LoopHeader = MI.getOperand(0).getMBB();
        auto *LoopEnd = &MBB;

        LLVM_DEBUG(dbgs() << "Checking Waterfall loop: "
                          << printMBBReference(*LoopHeader) << '\n');

        SmallSetVector<Register, 16> CandidateRegs;
        SmallVector<MachineInstr *, 16> Instructions;
        SmallSetVector<MachineBasicBlock *, 2> Blocks;

        collectWaterfallCandidateRegisters(LoopHeader, LoopEnd, CandidateRegs,
                                           Blocks, Instructions);
        MadeChange |= !CandidateRegs.empty();
        // Now we are safe to optimize.
        for (auto Reg : CandidateRegs)
          optimizeWaterfallLiveRange(Reg, LoopHeader, Blocks, Instructions);
      }
    }
  }

  return MadeChange;
}