//
// The LLVM Compiler Infrastructure
//
-// This file was developed by Bill Wendling and is distributed under the
-// University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass performs loop invariant code motion on machine instructions. We
// attempt to remove as much code from the body of a loop as possible.
//
+// This pass does not attempt to throttle itself to limit register pressure.
+// The register allocation phases are expected to perform rematerialization
+// to recover when register pressure is high.
+//
+// This pass is not intended to be a replacement or a complete alternative
+// for the LLVM-IR-level LICM pass. It is only designed to hoist simple
+// constructs that are not exposed before lowering and instruction selection.
+//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "machine-licm"
-#include "llvm/ADT/IndexedMap.h"
-#include "llvm/ADT/SmallVector.h"
-#include "llvm/ADT/Statistic.h"
-#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineDominators.h"
-#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
-#include "llvm/CodeGen/Passes.h"
+#include "llvm/CodeGen/MachineMemOperand.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/PseudoSourceValue.h"
+#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Target/TargetInstrInfo.h"
-#include "llvm/Support/CFG.h"
-#include "llvm/Support/CommandLine.h"
-#include "llvm/Support/Compiler.h"
-#include "llvm/Support/Debug.h"
-#include "llvm/Target/MRegisterInfo.h"
#include "llvm/Target/TargetMachine.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
using namespace llvm;
-namespace {
- // Hidden options to help debugging
- cl::opt<bool>
- PerformLICM("machine-licm",
- cl::init(false), cl::Hidden,
- cl::desc("Perform loop-invariant code motion on machine code"));
-}
-
STATISTIC(NumHoisted, "Number of machine instructions hoisted out of loops");
+STATISTIC(NumCSEed, "Number of hoisted machine instructions CSEed");
namespace {
- class VISIBILITY_HIDDEN MachineLICM : public MachineFunctionPass {
- // Various analyses that we use...
- MachineLoopInfo *LI; // Current MachineLoopInfo
- MachineDominatorTree *DT; // Machine dominator tree for the current Loop
-
+ class MachineLICM : public MachineFunctionPass {
+ MachineConstantPool *MCP;
+ const TargetMachine *TM;
const TargetInstrInfo *TII;
+ const TargetRegisterInfo *TRI;
+ BitVector AllocatableSet;
+
+ // Various analyses that we use...
+ AliasAnalysis *AA; // Alias analysis info.
+ MachineLoopInfo *LI; // Current MachineLoopInfo
+ MachineDominatorTree *DT; // Machine dominator tree for the cur loop
+ MachineRegisterInfo *RegInfo; // Machine register information
// State that is updated as we process loops
- bool Changed; // True if a loop is changed.
- MachineLoop *CurLoop; // The current loop we are working on.
+ bool Changed; // True if a loop is changed.
+ bool FirstInLoop; // True if it's the first LICM in the loop.
+ MachineLoop *CurLoop; // The current loop we are working on.
+ MachineBasicBlock *CurPreheader; // The preheader for CurLoop.
- // Map the def of a virtual register to the machine instruction.
- IndexedMap<const MachineInstr*, VirtReg2IndexFunctor> VRegDefs;
+ // For each opcode, keep a list of potentail CSE instructions.
+ DenseMap<unsigned, std::vector<const MachineInstr*> > CSEMap;
public:
static char ID; // Pass identification, replacement for typeid
- MachineLICM() : MachineFunctionPass((intptr_t)&ID) {}
+ MachineLICM() : MachineFunctionPass(&ID) {}
virtual bool runOnMachineFunction(MachineFunction &MF);
- /// FIXME: Loop preheaders?
- ///
+ const char *getPassName() const { return "Machine Instruction LICM"; }
+
+ // FIXME: Loop preheaders?
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesCFG();
AU.addRequired<MachineLoopInfo>();
AU.addRequired<MachineDominatorTree>();
- }
- private:
- /// VisitAllLoops - Visit all of the loops in depth first order and try to
- /// hoist invariant instructions from them.
- ///
- void VisitAllLoops(MachineLoop *L) {
- const std::vector<MachineLoop*> &SubLoops = L->getSubLoops();
-
- for (MachineLoop::iterator
- I = SubLoops.begin(), E = SubLoops.end(); I != E; ++I) {
- MachineLoop *ML = *I;
-
- // Traverse the body of the loop in depth first order on the dominator
- // tree so that we are guaranteed to see definitions before we see uses.
- VisitAllLoops(ML);
- HoistRegion(DT->getNode(ML->getHeader()));
- }
-
- HoistRegion(DT->getNode(L->getHeader()));
- }
-
- /// MapVirtualRegisterDefs - Create a map of which machine instruction
- /// defines a virtual register.
- ///
- void MapVirtualRegisterDefs(const MachineFunction &MF);
-
- /// IsInSubLoop - A little predicate that returns true if the specified
- /// basic block is in a subloop of the current one, not the current one
- /// itself.
- ///
- bool IsInSubLoop(MachineBasicBlock *BB) {
- assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
-
- for (MachineLoop::iterator
- I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I)
- if ((*I)->contains(BB))
- return true; // A subloop actually contains this block!
-
- return false;
+ AU.addRequired<AliasAnalysis>();
+ AU.addPreserved<MachineLoopInfo>();
+ AU.addPreserved<MachineDominatorTree>();
+ MachineFunctionPass::getAnalysisUsage(AU);
}
- /// CanHoistInst - Checks that this instructions is one that can be hoisted
- /// out of the loop. I.e., it has no side effects, isn't a control flow
- /// instr, etc.
- ///
- bool CanHoistInst(MachineInstr &I) const {
- const TargetInstrDescriptor *TID = I.getInstrDescriptor();
-
- // Don't hoist if this instruction implicitly reads physical registers or
- // doesn't take any operands.
- if (TID->ImplicitUses || !I.getNumOperands()) return false;
-
- MachineOpCode Opcode = TID->Opcode;
- return TII->isTriviallyReMaterializable(&I) &&
- // FIXME: Below necessary?
- !(TII->isReturn(Opcode) ||
- TII->isTerminatorInstr(Opcode) ||
- TII->isBranch(Opcode) ||
- TII->isIndirectBranch(Opcode) ||
- TII->isBarrier(Opcode) ||
- TII->isCall(Opcode) ||
- TII->isLoad(Opcode) || // TODO: Do loads and stores.
- TII->isStore(Opcode));
+ virtual void releaseMemory() {
+ CSEMap.clear();
}
+ private:
/// IsLoopInvariantInst - Returns true if the instruction is loop
/// invariant. I.e., all virtual register operands are defined outside of
/// the loop, physical registers aren't accessed (explicitly or implicitly),
///
bool IsLoopInvariantInst(MachineInstr &I);
- /// FindPredecessors - Get all of the predecessors of the loop that are not
- /// back-edges.
- ///
- void FindPredecessors(std::vector<MachineBasicBlock*> &Preds){
- const MachineBasicBlock *Header = CurLoop->getHeader();
-
- for (MachineBasicBlock::const_pred_iterator
- I = Header->pred_begin(), E = Header->pred_end(); I != E; ++I)
- if (!CurLoop->contains(*I))
- Preds.push_back(*I);
- }
-
- /// MoveInstToEndOfBlock - Moves the machine instruction to the bottom of
- /// the predecessor basic block (but before the terminator instructions).
- ///
- void MoveInstToEndOfBlock(MachineBasicBlock *MBB, MachineInstr *MI) {
- MachineBasicBlock::iterator Iter = MBB->getFirstTerminator();
- MBB->insert(Iter, MI);
- ++NumHoisted;
- }
+ /// IsProfitableToHoist - Return true if it is potentially profitable to
+ /// hoist the given loop invariant.
+ bool IsProfitableToHoist(MachineInstr &MI);
/// HoistRegion - Walk the specified region of the CFG (defined by all
/// blocks dominated by the specified block, and that are in the current
///
void HoistRegion(MachineDomTreeNode *N);
+ /// isLoadFromConstantMemory - Return true if the given instruction is a
+ /// load from constant memory.
+ bool isLoadFromConstantMemory(MachineInstr *MI);
+
+ /// ExtractHoistableLoad - Unfold a load from the given machineinstr if
+ /// the load itself could be hoisted. Return the unfolded and hoistable
+ /// load, or null if the load couldn't be unfolded or if it wouldn't
+ /// be hoistable.
+ MachineInstr *ExtractHoistableLoad(MachineInstr *MI);
+
+ /// LookForDuplicate - Find an instruction amount PrevMIs that is a
+ /// duplicate of MI. Return this instruction if it's found.
+ const MachineInstr *LookForDuplicate(const MachineInstr *MI,
+ std::vector<const MachineInstr*> &PrevMIs);
+
+ /// EliminateCSE - Given a LICM'ed instruction, look for an instruction on
+ /// the preheader that compute the same value. If it's found, do a RAU on
+ /// with the definition of the existing instruction rather than hoisting
+ /// the instruction to the preheader.
+ bool EliminateCSE(MachineInstr *MI,
+ DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator &CI);
+
/// Hoist - When an instruction is found to only use loop invariant operands
/// that is safe to hoist, this instruction is called to do the dirty work.
///
- void Hoist(MachineInstr &MI);
- };
+ void Hoist(MachineInstr *MI);
- char MachineLICM::ID = 0;
- RegisterPass<MachineLICM> X("machine-licm",
- "Machine Loop Invariant Code Motion");
+ /// InitCSEMap - Initialize the CSE map with instructions that are in the
+ /// current loop preheader that may become duplicates of instructions that
+ /// are hoisted out of the loop.
+ void InitCSEMap(MachineBasicBlock *BB);
+ };
} // end anonymous namespace
+char MachineLICM::ID = 0;
+static RegisterPass<MachineLICM>
+X("machinelicm", "Machine Loop Invariant Code Motion");
+
FunctionPass *llvm::createMachineLICMPass() { return new MachineLICM(); }
+/// LoopIsOuterMostWithPreheader - Test if the given loop is the outer-most
+/// loop that has a preheader.
+static bool LoopIsOuterMostWithPreheader(MachineLoop *CurLoop) {
+ for (MachineLoop *L = CurLoop->getParentLoop(); L; L = L->getParentLoop())
+ if (L->getLoopPreheader())
+ return false;
+ return true;
+}
+
/// Hoist expressions out of the specified loop. Note, alias info for inner loop
/// is not preserved so it is not a good idea to run LICM multiple times on one
/// loop.
///
bool MachineLICM::runOnMachineFunction(MachineFunction &MF) {
- if (!PerformLICM) return false; // For debugging.
+ DEBUG(dbgs() << "******** Machine LICM ********\n");
- Changed = false;
- TII = MF.getTarget().getInstrInfo();
+ Changed = FirstInLoop = false;
+ MCP = MF.getConstantPool();
+ TM = &MF.getTarget();
+ TII = TM->getInstrInfo();
+ TRI = TM->getRegisterInfo();
+ RegInfo = &MF.getRegInfo();
+ AllocatableSet = TRI->getAllocatableSet(MF);
// Get our Loop information...
LI = &getAnalysis<MachineLoopInfo>();
DT = &getAnalysis<MachineDominatorTree>();
+ AA = &getAnalysis<AliasAnalysis>();
- for (MachineLoopInfo::iterator
- I = LI->begin(), E = LI->end(); I != E; ++I) {
- MachineLoop *L = *I;
- CurLoop = L;
-
- // Visit all of the instructions of the loop. We want to visit the subloops
- // first, though, so that we can hoist their invariants first into their
- // containing loop before we process that loop.
- VisitAllLoops(L);
- }
-
- return Changed;
-}
-
-/// MapVirtualRegisterDefs - Create a map of which machine instruction defines a
-/// virtual register.
-///
-void MachineLICM::MapVirtualRegisterDefs(const MachineFunction &MF) {
- for (MachineFunction::const_iterator
- I = MF.begin(), E = MF.end(); I != E; ++I) {
- const MachineBasicBlock &MBB = *I;
+ for (MachineLoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) {
+ CurLoop = *I;
- for (MachineBasicBlock::const_iterator
- II = MBB.begin(), IE = MBB.end(); II != IE; ++II) {
- const MachineInstr &MI = *II;
+ // Only visit outer-most preheader-sporting loops.
+ if (!LoopIsOuterMostWithPreheader(CurLoop))
+ continue;
- for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
- const MachineOperand &MO = MI.getOperand(0);
+ // Determine the block to which to hoist instructions. If we can't find a
+ // suitable loop preheader, we can't do any hoisting.
+ //
+ // FIXME: We are only hoisting if the basic block coming into this loop
+ // has only one successor. This isn't the case in general because we haven't
+ // broken critical edges or added preheaders.
+ CurPreheader = CurLoop->getLoopPreheader();
+ if (!CurPreheader)
+ continue;
- if (MO.isRegister() && MO.isDef() &&
- MRegisterInfo::isVirtualRegister(MO.getReg()))
- VRegDefs[MO.getReg()] = &MI;
- }
- }
+ // CSEMap is initialized for loop header when the first instruction is
+ // being hoisted.
+ FirstInLoop = true;
+ HoistRegion(DT->getNode(CurLoop->getHeader()));
+ CSEMap.clear();
}
+
+ return Changed;
}
/// HoistRegion - Walk the specified region of the CFG (defined by all blocks
// If this subregion is not in the top level loop at all, exit.
if (!CurLoop->contains(BB)) return;
- // Only need to process the contents of this block if it is not part of a
- // subloop (which would already have been processed).
- if (!IsInSubLoop(BB))
- for (MachineBasicBlock::iterator
- I = BB->begin(), E = BB->end(); I != E; ) {
- MachineInstr &MI = *I++;
-
- // Try hoisting the instruction out of the loop. We can only do this if
- // all of the operands of the instruction are loop invariant and if it is
- // safe to hoist the instruction.
- Hoist(MI);
- }
+ for (MachineBasicBlock::iterator
+ MII = BB->begin(), E = BB->end(); MII != E; ) {
+ MachineBasicBlock::iterator NextMII = MII; ++NextMII;
+ Hoist(&*MII);
+ MII = NextMII;
+ }
const std::vector<MachineDomTreeNode*> &Children = N->getChildren();
/// IsLoopInvariantInst - Returns true if the instruction is loop
/// invariant. I.e., all virtual register operands are defined outside of the
-/// loop, physical registers aren't accessed (explicitly or implicitly), and the
-/// instruction is hoistable.
+/// loop, physical registers aren't accessed explicitly, and there are no side
+/// effects that aren't captured by the operands or other flags.
///
bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) {
- if (!CanHoistInst(I)) return false;
+ const TargetInstrDesc &TID = I.getDesc();
+
+ // Ignore stuff that we obviously can't hoist.
+ if (TID.mayStore() || TID.isCall() || TID.isTerminator() ||
+ TID.hasUnmodeledSideEffects())
+ return false;
+
+ if (TID.mayLoad()) {
+ // Okay, this instruction does a load. As a refinement, we allow the target
+ // to decide whether the loaded value is actually a constant. If so, we can
+ // actually use it as a load.
+ if (!I.isInvariantLoad(AA))
+ // FIXME: we should be able to hoist loads with no other side effects if
+ // there are no other instructions which can change memory in this loop.
+ // This is a trivial form of alias analysis.
+ return false;
+ }
+
+ DEBUG({
+ dbgs() << "--- Checking if we can hoist " << I;
+ if (I.getDesc().getImplicitUses()) {
+ dbgs() << " * Instruction has implicit uses:\n";
+
+ const TargetRegisterInfo *TRI = TM->getRegisterInfo();
+ for (const unsigned *ImpUses = I.getDesc().getImplicitUses();
+ *ImpUses; ++ImpUses)
+ dbgs() << " -> " << TRI->getName(*ImpUses) << "\n";
+ }
+
+ if (I.getDesc().getImplicitDefs()) {
+ dbgs() << " * Instruction has implicit defines:\n";
+
+ const TargetRegisterInfo *TRI = TM->getRegisterInfo();
+ for (const unsigned *ImpDefs = I.getDesc().getImplicitDefs();
+ *ImpDefs; ++ImpDefs)
+ dbgs() << " -> " << TRI->getName(*ImpDefs) << "\n";
+ }
+ });
- // The instruction is loop invariant if all of its operands are loop-invariant
+ if (I.getDesc().getImplicitDefs() || I.getDesc().getImplicitUses()) {
+ DEBUG(dbgs() << "Cannot hoist with implicit defines or uses\n");
+ return false;
+ }
+
+ // The instruction is loop invariant if all of its operands are.
for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
const MachineOperand &MO = I.getOperand(i);
- if (!MO.isRegister() || !MO.isUse())
+ if (!MO.isReg())
continue;
unsigned Reg = MO.getReg();
+ if (Reg == 0) continue;
+
+ // Don't hoist an instruction that uses or defines a physical register.
+ if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
+ if (MO.isUse()) {
+ // If the physreg has no defs anywhere, it's just an ambient register
+ // and we can freely move its uses. Alternatively, if it's allocatable,
+ // it could get allocated to something with a def during allocation.
+ if (!RegInfo->def_empty(Reg))
+ return false;
+ if (AllocatableSet.test(Reg))
+ return false;
+ // Check for a def among the register's aliases too.
+ for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
+ unsigned AliasReg = *Alias;
+ if (!RegInfo->def_empty(AliasReg))
+ return false;
+ if (AllocatableSet.test(AliasReg))
+ return false;
+ }
+ // Otherwise it's safe to move.
+ continue;
+ } else if (!MO.isDead()) {
+ // A def that isn't dead. We can't move it.
+ return false;
+ }
+ }
- // Don't hoist instructions that access physical registers.
- if (!MRegisterInfo::isVirtualRegister(Reg))
- return false;
+ if (!MO.isUse())
+ continue;
- assert(VRegDefs[Reg] && "Machine instr not mapped for this vreg?");
+ assert(RegInfo->getVRegDef(Reg) &&
+ "Machine instr not mapped for this vreg?!");
// If the loop contains the definition of an operand, then the instruction
// isn't loop invariant.
- if (CurLoop->contains(VRegDefs[Reg]->getParent()))
+ if (CurLoop->contains(RegInfo->getVRegDef(Reg)))
return false;
}
return true;
}
-/// Hoist - When an instruction is found to only use loop invariant operands
-/// that is safe to hoist, this instruction is called to do the dirty work.
-///
-void MachineLICM::Hoist(MachineInstr &MI) {
- if (!IsLoopInvariantInst(MI)) return;
- std::vector<MachineBasicBlock*> Preds;
+/// HasPHIUses - Return true if the specified register has any PHI use.
+static bool HasPHIUses(unsigned Reg, MachineRegisterInfo *RegInfo) {
+ for (MachineRegisterInfo::use_iterator UI = RegInfo->use_begin(Reg),
+ UE = RegInfo->use_end(); UI != UE; ++UI) {
+ MachineInstr *UseMI = &*UI;
+ if (UseMI->getOpcode() == TargetInstrInfo::PHI)
+ return true;
+ }
+ return false;
+}
+
+/// isLoadFromConstantMemory - Return true if the given instruction is a
+/// load from constant memory. Machine LICM will hoist these even if they are
+/// not re-materializable.
+bool MachineLICM::isLoadFromConstantMemory(MachineInstr *MI) {
+ if (!MI->getDesc().mayLoad()) return false;
+ if (!MI->hasOneMemOperand()) return false;
+ MachineMemOperand *MMO = *MI->memoperands_begin();
+ if (MMO->isVolatile()) return false;
+ if (!MMO->getValue()) return false;
+ const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(MMO->getValue());
+ if (PSV) {
+ MachineFunction &MF = *MI->getParent()->getParent();
+ return PSV->isConstant(MF.getFrameInfo());
+ } else {
+ return AA->pointsToConstantMemory(MMO->getValue());
+ }
+}
+
+/// IsProfitableToHoist - Return true if it is potentially profitable to hoist
+/// the given loop invariant.
+bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) {
+ if (MI.getOpcode() == TargetInstrInfo::IMPLICIT_DEF)
+ return false;
+
+ // FIXME: For now, only hoist re-materilizable instructions. LICM will
+ // increase register pressure. We want to make sure it doesn't increase
+ // spilling.
+ // Also hoist loads from constant memory, e.g. load from stubs, GOT. Hoisting
+ // these tend to help performance in low register pressure situation. The
+ // trade off is it may cause spill in high pressure situation. It will end up
+ // adding a store in the loop preheader. But the reload is no more expensive.
+ // The side benefit is these loads are frequently CSE'ed.
+ if (!TII->isTriviallyReMaterializable(&MI, AA)) {
+ if (!isLoadFromConstantMemory(&MI))
+ return false;
+ }
+
+ // If result(s) of this instruction is used by PHIs, then don't hoist it.
+ // The presence of joins makes it difficult for current register allocator
+ // implementation to perform remat.
+ for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = MI.getOperand(i);
+ if (!MO.isReg() || !MO.isDef())
+ continue;
+ if (HasPHIUses(MO.getReg(), RegInfo))
+ return false;
+ }
+
+ return true;
+}
+
+MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) {
+ // If not, we may be able to unfold a load and hoist that.
+ // First test whether the instruction is loading from an amenable
+ // memory location.
+ if (!isLoadFromConstantMemory(MI))
+ return 0;
+
+ // Next determine the register class for a temporary register.
+ unsigned LoadRegIndex;
+ unsigned NewOpc =
+ TII->getOpcodeAfterMemoryUnfold(MI->getOpcode(),
+ /*UnfoldLoad=*/true,
+ /*UnfoldStore=*/false,
+ &LoadRegIndex);
+ if (NewOpc == 0) return 0;
+ const TargetInstrDesc &TID = TII->get(NewOpc);
+ if (TID.getNumDefs() != 1) return 0;
+ const TargetRegisterClass *RC = TID.OpInfo[LoadRegIndex].getRegClass(TRI);
+ // Ok, we're unfolding. Create a temporary register and do the unfold.
+ unsigned Reg = RegInfo->createVirtualRegister(RC);
+
+ MachineFunction &MF = *MI->getParent()->getParent();
+ SmallVector<MachineInstr *, 2> NewMIs;
+ bool Success =
+ TII->unfoldMemoryOperand(MF, MI, Reg,
+ /*UnfoldLoad=*/true, /*UnfoldStore=*/false,
+ NewMIs);
+ (void)Success;
+ assert(Success &&
+ "unfoldMemoryOperand failed when getOpcodeAfterMemoryUnfold "
+ "succeeded!");
+ assert(NewMIs.size() == 2 &&
+ "Unfolded a load into multiple instructions!");
+ MachineBasicBlock *MBB = MI->getParent();
+ MBB->insert(MI, NewMIs[0]);
+ MBB->insert(MI, NewMIs[1]);
+ // If unfolding produced a load that wasn't loop-invariant or profitable to
+ // hoist, discard the new instructions and bail.
+ if (!IsLoopInvariantInst(*NewMIs[0]) || !IsProfitableToHoist(*NewMIs[0])) {
+ NewMIs[0]->eraseFromParent();
+ NewMIs[1]->eraseFromParent();
+ return 0;
+ }
+ // Otherwise we successfully unfolded a load that we can hoist.
+ MI->eraseFromParent();
+ return NewMIs[0];
+}
- // Non-back-edge predecessors.
- FindPredecessors(Preds);
+void MachineLICM::InitCSEMap(MachineBasicBlock *BB) {
+ for (MachineBasicBlock::iterator I = BB->begin(),E = BB->end(); I != E; ++I) {
+ const MachineInstr *MI = &*I;
+ // FIXME: For now, only hoist re-materilizable instructions. LICM will
+ // increase register pressure. We want to make sure it doesn't increase
+ // spilling.
+ if (TII->isTriviallyReMaterializable(MI, AA)) {
+ unsigned Opcode = MI->getOpcode();
+ DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator
+ CI = CSEMap.find(Opcode);
+ if (CI != CSEMap.end())
+ CI->second.push_back(MI);
+ else {
+ std::vector<const MachineInstr*> CSEMIs;
+ CSEMIs.push_back(MI);
+ CSEMap.insert(std::make_pair(Opcode, CSEMIs));
+ }
+ }
+ }
+}
- // Either we don't have any predecessors(?!) or we have more than one, which
- // is forbidden.
- if (Preds.empty() || Preds.size() != 1) return;
+const MachineInstr*
+MachineLICM::LookForDuplicate(const MachineInstr *MI,
+ std::vector<const MachineInstr*> &PrevMIs) {
+ for (unsigned i = 0, e = PrevMIs.size(); i != e; ++i) {
+ const MachineInstr *PrevMI = PrevMIs[i];
+ if (TII->isIdentical(MI, PrevMI, RegInfo))
+ return PrevMI;
+ }
+ return 0;
+}
- // Check that the predecessor is qualified to take the hoisted
- // instruction. I.e., there is only one edge from the predecessor, and it's to
- // the loop header.
- MachineBasicBlock *MBB = Preds.front();
+bool MachineLICM::EliminateCSE(MachineInstr *MI,
+ DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator &CI) {
+ if (CI == CSEMap.end())
+ return false;
+
+ if (const MachineInstr *Dup = LookForDuplicate(MI, CI->second)) {
+ DEBUG(dbgs() << "CSEing " << *MI << " with " << *Dup);
+ for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = MI->getOperand(i);
+ if (MO.isReg() && MO.isDef())
+ RegInfo->replaceRegWith(MO.getReg(), Dup->getOperand(i).getReg());
+ }
+ MI->eraseFromParent();
+ ++NumCSEed;
+ return true;
+ }
+ return false;
+}
- // FIXME: We are assuming at first that the basic block coming into this loop
- // has only one successor. This isn't the case in general because we haven't
- // broken critical edges or added preheaders.
- if (MBB->succ_size() != 1) return;
- assert(*MBB->succ_begin() == CurLoop->getHeader() &&
- "The predecessor doesn't feed directly into the loop header!");
+/// Hoist - When an instruction is found to use only loop invariant operands
+/// that are safe to hoist, this instruction is called to do the dirty work.
+///
+void MachineLICM::Hoist(MachineInstr *MI) {
+ // First check whether we should hoist this instruction.
+ if (!IsLoopInvariantInst(*MI) || !IsProfitableToHoist(*MI)) {
+ // If not, try unfolding a hoistable load.
+ MI = ExtractHoistableLoad(MI);
+ if (!MI) return;
+ }
- // Now move the instructions to the predecessor.
- MoveInstToEndOfBlock(MBB, MI.clone());
+ // Now move the instructions to the predecessor, inserting it before any
+ // terminator instructions.
+ DEBUG({
+ dbgs() << "Hoisting " << *MI;
+ if (CurPreheader->getBasicBlock())
+ dbgs() << " to MachineBasicBlock "
+ << CurPreheader->getName();
+ if (MI->getParent()->getBasicBlock())
+ dbgs() << " from MachineBasicBlock "
+ << MI->getParent()->getName();
+ dbgs() << "\n";
+ });
+
+ // If this is the first instruction being hoisted to the preheader,
+ // initialize the CSE map with potential common expressions.
+ InitCSEMap(CurPreheader);
+
+ // Look for opportunity to CSE the hoisted instruction.
+ unsigned Opcode = MI->getOpcode();
+ DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator
+ CI = CSEMap.find(Opcode);
+ if (!EliminateCSE(MI, CI)) {
+ // Otherwise, splice the instruction to the preheader.
+ CurPreheader->splice(CurPreheader->getFirstTerminator(),MI->getParent(),MI);
+
+ // Add to the CSE map.
+ if (CI != CSEMap.end())
+ CI->second.push_back(MI);
+ else {
+ std::vector<const MachineInstr*> CSEMIs;
+ CSEMIs.push_back(MI);
+ CSEMap.insert(std::make_pair(Opcode, CSEMIs));
+ }
+ }
- // Hoisting was successful! Remove bothersome instruction now.
- MI.getParent()->remove(&MI);
+ ++NumHoisted;
Changed = true;
}