R600: Add a pass that merge Vector Register

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@183279 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/R600/AMDGPU.h

@@ -23,6 +23,7 @@ class AMDGPUTargetMachine;
 // R600 Passes
 FunctionPass* createR600TextureIntrinsicsReplacer();
 FunctionPass* createR600KernelParametersPass(const DataLayout *TD);
+FunctionPass *createR600VectorRegMerger(TargetMachine &tm);
 FunctionPass *createR600ExpandSpecialInstrsPass(TargetMachine &tm);
 FunctionPass *createR600EmitClauseMarkers(TargetMachine &tm);
 FunctionPass *createR600Packetizer(TargetMachine &tm);

lib/Target/R600/AMDGPUTargetMachine.cpp

@@ -130,6 +130,11 @@ bool AMDGPUPassConfig::addInstSelector() {
 
 bool AMDGPUPassConfig::addPreRegAlloc() {
   addPass(createAMDGPUConvertToISAPass(*TM));
+  const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
+  
+  if (ST.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX) {
+    addPass(createR600VectorRegMerger(*TM));
+  }
   return false;
 }
 

lib/Target/R600/CMakeLists.txt

@@ -41,6 +41,7 @@ add_llvm_target(R600CodeGen
   R600ISelLowering.cpp
   R600MachineFunctionInfo.cpp
   R600MachineScheduler.cpp
+  R600OptimizeVectorRegisters.cpp
   R600Packetizer.cpp
   R600RegisterInfo.cpp
   R600TextureIntrinsicsReplacer.cpp

lib/Target/R600/R600OptimizeVectorRegisters.cpp

@@ -0,0 +1,363 @@
+//===--------------------- R600MergeVectorRegisters.cpp -------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// This pass merges inputs of swizzeable instructions into vector sharing
+/// common data and/or have enough undef subreg using swizzle abilities.
+///
+/// For instance let's consider the following pseudo code :
+/// vreg5<def> = REG_SEQ vreg1, sub0, vreg2, sub1, vreg3, sub2, undef, sub3
+/// ...
+/// vreg7<def> = REG_SEQ vreg1, sub0, vreg3, sub1, undef, sub2, vreg4, sub3
+/// (swizzable Inst) vreg7, SwizzleMask : sub0, sub1, sub2, sub3
+///
+/// is turned into :
+/// vreg5<def> = REG_SEQ vreg1, sub0, vreg2, sub1, vreg3, sub2, undef, sub3
+/// ...
+/// vreg7<def> = INSERT_SUBREG vreg4, sub3
+/// (swizzable Inst) vreg7, SwizzleMask : sub0, sub2, sub1, sub3
+///
+/// This allow regalloc to reduce register pressure for vector registers and
+/// to reduce MOV count.
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "vec-merger"
+#include "llvm/Support/Debug.h"
+#include "AMDGPU.h"
+#include "R600InstrInfo.h"
+#include "llvm/CodeGen/DFAPacketizer.h"
+#include "llvm/CodeGen/MachineDominators.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+
+using namespace llvm;
+
+namespace {
+
+static bool
+isImplicitlyDef(MachineRegisterInfo &MRI, unsigned Reg) {
+  for (MachineRegisterInfo::def_iterator It = MRI.def_begin(Reg),
+      E = MRI.def_end(); It != E; ++It) {
+    return (*It).isImplicitDef();
+  }
+  llvm_unreachable("Reg without a def");
+  return false;
+}
+
+class RegSeqInfo {
+public:
+  MachineInstr *Instr;
+  DenseMap<unsigned, unsigned> RegToChan;
+  std::vector<unsigned> UndefReg;
+  RegSeqInfo(MachineRegisterInfo &MRI, MachineInstr *MI) : Instr(MI) {
+    assert (MI->getOpcode() == AMDGPU::REG_SEQUENCE);
+    for (unsigned i = 1, e = Instr->getNumOperands(); i < e; i+=2) {
+      MachineOperand &MO = Instr->getOperand(i);
+      unsigned Chan = Instr->getOperand(i + 1).getImm();
+      if (isImplicitlyDef(MRI, MO.getReg()))
+        UndefReg.push_back(Chan);
+      else
+        RegToChan[MO.getReg()] = Chan;
+    }
+  }
+  RegSeqInfo() {}
+
+  bool operator==(const RegSeqInfo &RSI) const {
+    return RSI.Instr == Instr;
+  }
+};
+
+class R600VectorRegMerger : public MachineFunctionPass {
+private:
+  MachineRegisterInfo *MRI;
+  const R600InstrInfo *TII;
+  bool canSwizzle(const MachineInstr &) const;
+  bool areAllUsesSwizzeable(unsigned Reg) const;
+  void SwizzleInput(MachineInstr &,
+      const std::vector<std::pair<unsigned, unsigned> > &) const;
+  bool tryMergeVector(const RegSeqInfo *, RegSeqInfo *,
+      std::vector<std::pair<unsigned, unsigned> > &Remap) const;
+  bool tryMergeUsingCommonSlot(RegSeqInfo &RSI, RegSeqInfo &CompatibleRSI,
+      std::vector<std::pair<unsigned, unsigned> > &RemapChan);
+  bool tryMergeUsingFreeSlot(RegSeqInfo &RSI, RegSeqInfo &CompatibleRSI,
+      std::vector<std::pair<unsigned, unsigned> > &RemapChan);
+  MachineInstr *RebuildVector(RegSeqInfo *MI,
+      const RegSeqInfo *BaseVec,
+      const std::vector<std::pair<unsigned, unsigned> > &RemapChan) const;
+  void RemoveMI(MachineInstr *);
+  void trackRSI(const RegSeqInfo &RSI);
+
+  typedef DenseMap<unsigned, std::vector<MachineInstr *> > InstructionSetMap;
+  DenseMap<MachineInstr *, RegSeqInfo> PreviousRegSeq;
+  InstructionSetMap PreviousRegSeqByReg;
+  InstructionSetMap PreviousRegSeqByUndefCount;
+public:
+  static char ID;
+  R600VectorRegMerger(TargetMachine &tm) : MachineFunctionPass(ID),
+  TII (static_cast<const R600InstrInfo *>(tm.getInstrInfo())) { }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const {
+    AU.setPreservesCFG();
+    AU.addRequired<MachineDominatorTree>();
+    AU.addPreserved<MachineDominatorTree>();
+    AU.addRequired<MachineLoopInfo>();
+    AU.addPreserved<MachineLoopInfo>();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+
+  const char *getPassName() const {
+    return "R600 Vector Registers Merge Pass";
+  }
+
+  bool runOnMachineFunction(MachineFunction &Fn);
+};
+
+char R600VectorRegMerger::ID = 0;
+
+bool R600VectorRegMerger::canSwizzle(const MachineInstr &MI)
+    const {
+  if (TII->get(MI.getOpcode()).TSFlags & R600_InstFlag::TEX_INST)
+    return true;
+  switch (MI.getOpcode()) {
+  case AMDGPU::R600_ExportSwz:
+  case AMDGPU::EG_ExportSwz:
+    return true;
+  default:
+    return false;
+  }
+}
+
+bool R600VectorRegMerger::tryMergeVector(const RegSeqInfo *Untouched,
+    RegSeqInfo *ToMerge, std::vector< std::pair<unsigned, unsigned> > &Remap)
+    const {
+  unsigned CurrentUndexIdx = 0;
+  for (DenseMap<unsigned, unsigned>::iterator It = ToMerge->RegToChan.begin(),
+      E = ToMerge->RegToChan.end(); It != E; ++It) {
+    DenseMap<unsigned, unsigned>::const_iterator PosInUntouched =
+        Untouched->RegToChan.find((*It).first);
+    if (PosInUntouched != Untouched->RegToChan.end()) {
+      Remap.push_back(std::pair<unsigned, unsigned>
+          ((*It).second, (*PosInUntouched).second));
+      continue;
+    }
+    if (CurrentUndexIdx >= Untouched->UndefReg.size())
+      return false;
+    Remap.push_back(std::pair<unsigned, unsigned>
+        ((*It).second, Untouched->UndefReg[CurrentUndexIdx++]));
+  }
+
+  return true;
+}
+
+MachineInstr *R600VectorRegMerger::RebuildVector(
+    RegSeqInfo *RSI, const RegSeqInfo *BaseRSI,
+    const std::vector<std::pair<unsigned, unsigned> > &RemapChan) const {
+  unsigned Reg = RSI->Instr->getOperand(0).getReg();
+  MachineBasicBlock::iterator Pos = RSI->Instr;
+  MachineBasicBlock &MBB = *Pos->getParent();
+  DebugLoc DL = Pos->getDebugLoc();
+
+  unsigned SrcVec = BaseRSI->Instr->getOperand(0).getReg();
+  DenseMap<unsigned, unsigned> UpdatedRegToChan = BaseRSI->RegToChan;
+  std::vector<unsigned> UpdatedUndef = BaseRSI->UndefReg;
+  for (DenseMap<unsigned, unsigned>::iterator It = RSI->RegToChan.begin(),
+      E = RSI->RegToChan.end(); It != E; ++It) {
+    if (BaseRSI->RegToChan.find((*It).first) != BaseRSI->RegToChan.end()) {
+      UpdatedRegToChan[(*It).first] = (*It).second;
+      continue;
+    }
+    unsigned DstReg = MRI->createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
+    unsigned SubReg = (*It).first;
+    unsigned Swizzle = (*It).second;
+    unsigned Chan;
+    for (unsigned j = 0, je = RemapChan.size(); j < je; j++) {
+      if (RemapChan[j].first == Swizzle) {
+        Chan = RemapChan[j].second;
+        break;
+      }
+    }
+    MachineInstr *Tmp = BuildMI(MBB, Pos, DL, TII->get(AMDGPU::INSERT_SUBREG),
+        DstReg)
+        .addReg(SrcVec)
+        .addReg(SubReg)
+        .addImm(Chan);
+    UpdatedRegToChan[SubReg] = Chan;
+    for (std::vector<unsigned>::iterator RemoveIt = UpdatedUndef.begin(),
+        RemoveE = UpdatedUndef.end(); RemoveIt != RemoveE; ++ RemoveIt) {
+      if (*RemoveIt == Chan)
+        UpdatedUndef.erase(RemoveIt);
+    }
+    DEBUG(dbgs() << "    ->"; Tmp->dump(););
+    SrcVec = DstReg;
+  }
+  Pos = BuildMI(MBB, Pos, DL, TII->get(AMDGPU::COPY), Reg)
+      .addReg(SrcVec);
+  DEBUG(dbgs() << "    ->"; Pos->dump(););
+
+  DEBUG(dbgs() << "  Updating Swizzle:\n");
+  for (MachineRegisterInfo::use_iterator It = MRI->use_begin(Reg),
+      E = MRI->use_end(); It != E; ++It) {
+    DEBUG(dbgs() << "    ";(*It).dump(); dbgs() << "    ->");
+    SwizzleInput(*It, RemapChan);
+    DEBUG((*It).dump());
+  }
+  RSI->Instr->eraseFromParent();
+
+  // Update RSI
+  RSI->Instr = Pos;
+  RSI->RegToChan = UpdatedRegToChan;
+  RSI->UndefReg = UpdatedUndef;
+
+  return Pos;
+}
+
+void R600VectorRegMerger::RemoveMI(MachineInstr *MI) {
+  for (InstructionSetMap::iterator It = PreviousRegSeqByReg.begin(),
+      E = PreviousRegSeqByReg.end(); It != E; ++It) {
+    std::vector<MachineInstr *> &MIs = (*It).second;
+    MIs.erase(std::find(MIs.begin(), MIs.end(), MI), MIs.end());
+  }
+  for (InstructionSetMap::iterator It = PreviousRegSeqByUndefCount.begin(),
+      E = PreviousRegSeqByUndefCount.end(); It != E; ++It) {
+    std::vector<MachineInstr *> &MIs = (*It).second;
+    MIs.erase(std::find(MIs.begin(), MIs.end(), MI), MIs.end());
+  }
+}
+
+void R600VectorRegMerger::SwizzleInput(MachineInstr &MI,
+    const std::vector<std::pair<unsigned, unsigned> > &RemapChan) const {
+  unsigned Offset;
+  if (TII->get(MI.getOpcode()).TSFlags & R600_InstFlag::TEX_INST)
+    Offset = 2;
+  else
+    Offset = 3;
+  for (unsigned i = 0; i < 4; i++) {
+    unsigned Swizzle = MI.getOperand(i + Offset).getImm() + 1;
+    for (unsigned j = 0, e = RemapChan.size(); j < e; j++) {
+      if (RemapChan[j].first == Swizzle) {
+        MI.getOperand(i + Offset).setImm(RemapChan[j].second - 1);
+        break;
+      }
+    }
+  }
+}
+
+bool R600VectorRegMerger::areAllUsesSwizzeable(unsigned Reg) const {
+  for (MachineRegisterInfo::use_iterator It = MRI->use_begin(Reg),
+      E = MRI->use_end(); It != E; ++It) {
+    if (!canSwizzle(*It))
+      return false;
+  }
+  return true;
+}
+
+bool R600VectorRegMerger::tryMergeUsingCommonSlot(RegSeqInfo &RSI,
+    RegSeqInfo &CompatibleRSI,
+    std::vector<std::pair<unsigned, unsigned> > &RemapChan) {
+  for (MachineInstr::mop_iterator MOp = RSI.Instr->operands_begin(),
+      MOE = RSI.Instr->operands_end(); MOp != MOE; ++MOp) {
+    if (!MOp->isReg())
+      continue;
+    if (PreviousRegSeqByReg[MOp->getReg()].empty())
+      continue;
+    std::vector<MachineInstr *> MIs = PreviousRegSeqByReg[MOp->getReg()];
+    for (unsigned i = 0, e = MIs.size(); i < e; i++) {
+      CompatibleRSI = PreviousRegSeq[MIs[i]];
+      if (RSI == CompatibleRSI)
+        continue;
+      if (tryMergeVector(&CompatibleRSI, &RSI, RemapChan))
+        return true;
+    }
+  }
+  return false;
+}
+
+bool R600VectorRegMerger::tryMergeUsingFreeSlot(RegSeqInfo &RSI,
+    RegSeqInfo &CompatibleRSI,
+    std::vector<std::pair<unsigned, unsigned> > &RemapChan) {
+  unsigned NeededUndefs = 4 - RSI.UndefReg.size();
+  if (PreviousRegSeqByUndefCount[NeededUndefs].empty())
+    return false;
+  std::vector<MachineInstr *> &MIs =
+      PreviousRegSeqByUndefCount[NeededUndefs];
+  CompatibleRSI = PreviousRegSeq[MIs.back()];
+  tryMergeVector(&CompatibleRSI, &RSI, RemapChan);
+  return true;
+}
+
+void R600VectorRegMerger::trackRSI(const RegSeqInfo &RSI) {
+  for (DenseMap<unsigned, unsigned>::const_iterator
+  It = RSI.RegToChan.begin(), E = RSI.RegToChan.end(); It != E; ++It) {
+    PreviousRegSeqByReg[(*It).first].push_back(RSI.Instr);
+  }
+  PreviousRegSeqByUndefCount[RSI.UndefReg.size()].push_back(RSI.Instr);
+  PreviousRegSeq[RSI.Instr] = RSI;
+}
+
+bool R600VectorRegMerger::runOnMachineFunction(MachineFunction &Fn) {
+  MRI = &(Fn.getRegInfo());
+  for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
+       MBB != MBBe; ++MBB) {
+    MachineBasicBlock *MB = MBB;
+    PreviousRegSeq.clear();
+    PreviousRegSeqByReg.clear();
+    PreviousRegSeqByUndefCount.clear();
+
+    for (MachineBasicBlock::iterator MII = MB->begin(), MIIE = MB->end();
+         MII != MIIE; ++MII) {
+      MachineInstr *MI = MII;
+      if (MI->getOpcode() != AMDGPU::REG_SEQUENCE)
+        continue;
+
+      RegSeqInfo RSI(*MRI, MI);
+
+      // All uses of MI are swizzeable ?
+      unsigned Reg = MI->getOperand(0).getReg();
+      if (!areAllUsesSwizzeable(Reg))
+        continue;
+
+      DEBUG (dbgs() << "Trying to optimize ";
+          MI->dump();
+      );
+
+      RegSeqInfo CandidateRSI;
+      std::vector<std::pair<unsigned, unsigned> > RemapChan;
+      DEBUG(dbgs() << "Using common slots...\n";);
+      if (tryMergeUsingCommonSlot(RSI, CandidateRSI, RemapChan)) {
+        // Remove CandidateRSI mapping
+        RemoveMI(CandidateRSI.Instr);
+        MII = RebuildVector(&RSI, &CandidateRSI, RemapChan);
+        trackRSI(RSI);
+        continue;
+      }
+      DEBUG(dbgs() << "Using free slots...\n";);
+      RemapChan.clear();
+      if (tryMergeUsingFreeSlot(RSI, CandidateRSI, RemapChan)) {
+        RemoveMI(CandidateRSI.Instr);
+        MII = RebuildVector(&RSI, &CandidateRSI, RemapChan);
+        trackRSI(RSI);
+        continue;
+      }
+      //Failed to merge
+      trackRSI(RSI);
+    }
+  }
+  return false;
+}
+
+}
+
+llvm::FunctionPass *llvm::createR600VectorRegMerger(TargetMachine &tm) {
+  return new R600VectorRegMerger(tm);
+}
+

test/CodeGen/R600/texture-input-merge.ll

@@ -0,0 +1,30 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK-NOT: MOV
+
+define void @test() {
+  %1 = call float @llvm.R600.load.input(i32 0)
+  %2 = call float @llvm.R600.load.input(i32 1)
+  %3 = call float @llvm.R600.load.input(i32 2)
+  %4 = call float @llvm.R600.load.input(i32 3)
+  %5 = fmul float %1, 3.0
+  %6 = fmul float %2, 3.0
+  %7 = fmul float %3, 3.0
+  %8 = fmul float %4, 3.0
+  %9 = insertelement <4 x float> undef, float %5, i32 0
+  %10 = insertelement <4 x float> %9, float %6, i32 1
+  %11 = insertelement <4 x float> undef, float %7, i32 0
+  %12 = insertelement <4 x float> %11, float %5, i32 1
+  %13 = insertelement <4 x float> undef, float %8, i32 0
+  %14 = call <4 x float> @llvm.R600.tex(<4 x float> %10, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0)
+  %15 = call <4 x float> @llvm.R600.tex(<4 x float> %12, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0)
+  %16 = call <4 x float> @llvm.R600.tex(<4 x float> %13, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0)
+  %17 = fadd <4 x float> %14, %15
+  %18 = fadd <4 x float> %17, %16
+  call void @llvm.R600.store.swizzle(<4 x float> %18, i32 0, i32 0)
+  ret void
+}
+
+declare float @llvm.R600.load.input(i32) readnone
+declare <4 x float> @llvm.R600.tex(<4 x float>, i32, i32, i32, i32, i32, i32, i32, i32, i32) readnone
+declare void @llvm.R600.store.swizzle(<4 x float>, i32, i32)