Add a late optimization to BranchFolding that hoist common instruction sequences

at the start of basic blocks to their common predecessor. It's actually quite common (e.g. about 50 times in JM/lencod) and has shown to be a nice code size benefit. e.g. pushq %rax testl %edi, %edi jne LBB0_2 ## BB#1: xorb %al, %al popq %rdx ret LBB0_2: xorb %al, %al callq _foo popq %rdx ret => pushq %rax xorb %al, %al testl %edi, %edi je LBB0_2 ## BB#1: callq _foo LBB0_2: popq %rdx ret rdar://9145558 llvm-svn: 131172
2011-05-11 01:03:01 +00:00 · 2011-05-11 01:03:01 +00:00 · 05fc35e275
parent 4a34e61e53
commit 05fc35e275
4 changed files with 296 additions and 10 deletions
--- a/llvm/lib/CodeGen/BranchFolding.cpp
+++ b/llvm/lib/CodeGen/BranchFolding.cpp
@ -41,6 +41,7 @@ using namespace llvm;
 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
 STATISTIC(NumBranchOpts, "Number of branches optimized");
 STATISTIC(NumTailMerge , "Number of block tails merged");
+STATISTIC(NumHoist     , "Number of times common instructions are hoisted");

 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
                              cl::init(cl::BOU_UNSET), cl::Hidden);
@ -65,7 +66,7 @@ namespace {
  public:
    static char ID;
    explicit BranchFolderPass(bool defaultEnableTailMerge)
-      : MachineFunctionPass(ID), BranchFolder(defaultEnableTailMerge) {}
+      : MachineFunctionPass(ID), BranchFolder(defaultEnableTailMerge, true) {}

    virtual bool runOnMachineFunction(MachineFunction &MF);
    virtual const char *getPassName() const { return "Control Flow Optimizer"; }
@ -86,12 +87,14 @@ bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
 }


-BranchFolder::BranchFolder(bool defaultEnableTailMerge) {
+BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist) {
  switch (FlagEnableTailMerge) {
  case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
  case cl::BOU_TRUE: EnableTailMerge = true; break;
  case cl::BOU_FALSE: EnableTailMerge = false; break;
  }
+
+  EnableHoistCommonCode = CommonHoist;
 }

 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
@ -186,9 +189,10 @@ bool BranchFolder::OptimizeFunction(MachineFunction &MF,

  bool MadeChangeThisIteration = true;
  while (MadeChangeThisIteration) {
-    MadeChangeThisIteration = false;
-    MadeChangeThisIteration |= TailMergeBlocks(MF);
+    MadeChangeThisIteration    = TailMergeBlocks(MF);
    MadeChangeThisIteration   |= OptimizeBranches(MF);
+    if (EnableHoistCommonCode)
+      MadeChangeThisIteration |= HoistCommonCode(MF);
    MadeChange |= MadeChangeThisIteration;
  }

@ -910,7 +914,8 @@ bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
  // Make sure blocks are numbered in order
  MF.RenumberBlocks();

-  for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
+  for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
+       I != E; ) {
    MachineBasicBlock *MBB = I++;
    MadeChange |= OptimizeBlock(MBB);

@ -1339,3 +1344,253 @@ ReoptimizeBlock:

  return MadeChange;
 }
+
+//===----------------------------------------------------------------------===//
+//  Hoist Common Code
+//===----------------------------------------------------------------------===//
+
+/// HoistCommonCode - Hoist common instruction sequences at the start of basic
+/// blocks to their common predecessor.
+/// NOTE: This optimization does not update live-in information so it must be
+/// run after all passes that require correct liveness information.
+bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
+  bool MadeChange = false;
+  for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
+    MachineBasicBlock *MBB = I++;
+    MadeChange |= HoistCommonCodeInSuccs(MBB);
+  }
+
+  return MadeChange;
+}
+
+/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
+/// its 'true' successor.
+static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
+                                         MachineBasicBlock *TrueBB) {
+  for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
+         E = BB->succ_end(); SI != E; ++SI) {
+    MachineBasicBlock *SuccBB = *SI;
+    if (SuccBB != TrueBB)
+      return SuccBB;
+  }
+  return NULL;
+}
+
+/// findHoistingInsertPosAndDeps - Find the location to move common instructions
+/// in successors to. The location is ususally just before the terminator,
+/// however if the terminator is a conditional branch and its previous
+/// instruction is the flag setting instruction, the previous instruction is
+/// the preferred location. This function also gathers uses and defs of the
+/// instructions from the insertion point to the end of the block. The data is
+/// used by HoistCommonCodeInSuccs to ensure safety.
+static
+MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
+                                                  const TargetInstrInfo *TII,
+                                                  const TargetRegisterInfo *TRI,
+                                                  SmallSet<unsigned,4> &Uses,
+                                                  SmallSet<unsigned,4> &Defs) {
+  MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
+  if (!TII->isUnpredicatedTerminator(Loc))
+    return MBB->end();
+
+  for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
+    const MachineOperand &MO = Loc->getOperand(i);
+    if (!MO.isReg())
+      continue;
+    unsigned Reg = MO.getReg();
+    if (!Reg)
+      continue;
+    if (MO.isUse()) {
+      Uses.insert(Reg);
+      for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
+        Uses.insert(*AS);
+    } else if (!MO.isDead())
+      // Don't try to hoist code in the rare case the terminator defines a
+      // register that is later used.
+      return MBB->end();
+  }
+
+  if (Uses.empty())
+    return Loc;
+  if (Loc == MBB->begin())
+    return MBB->end();
+
+  // The terminator is probably a conditional branch, try not to separate the
+  // branch from condition setting instruction.
+  MachineBasicBlock::iterator PI = Loc;
+  --PI;
+  while (PI != MBB->begin() && Loc->isDebugValue())
+    --PI;
+
+  bool IsDef = false;
+  for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
+    const MachineOperand &MO = PI->getOperand(i);
+    if (!MO.isReg() || MO.isUse())
+      continue;
+    unsigned Reg = MO.getReg();
+    if (!Reg)
+      continue;
+    if (Uses.count(Reg))
+      IsDef = true;
+  }
+  if (!IsDef)
+    // The condition setting instruction is not just before the conditional
+    // branch.
+    return Loc;
+
+  // Be conservative, don't insert instruction above something that may have
+  // side-effects. And since it's potentially bad to separate flag setting
+  // instruction from the conditional branch, just abort the optimization
+  // completely.
+  // Also avoid moving code above predicated instruction since it's hard to
+  // reason about register liveness with predicated instruction.
+  bool DontMoveAcrossStore = true;
+  if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) ||
+      TII->isPredicated(PI))
+    return MBB->end();
+
+
+  // Find out what registers are live. Note this routine is ignoring other live
+  // registers which are only used by instructions in successor blocks.
+  for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
+    const MachineOperand &MO = PI->getOperand(i);
+    if (!MO.isReg())
+      continue;
+    unsigned Reg = MO.getReg();
+    if (!Reg)
+      continue;
+    if (MO.isUse()) {
+      Uses.insert(Reg);
+      for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
+        Uses.insert(*AS);
+    } else {
+      if (Uses.count(Reg)) {
+        Uses.erase(Reg);
+        for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
+          Uses.erase(*SR); // Use getSubRegisters to be conservative
+        Defs.insert(Reg);
+        for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
+          Defs.insert(*AS);
+      }
+    }
+  }
+
+  return PI;
+}
+
+/// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
+/// sequence at the start of the function, move the instructions before MBB
+/// terminator if it's legal.
+bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
+  MachineBasicBlock *TBB = 0, *FBB = 0;
+  SmallVector<MachineOperand, 4> Cond;
+  if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
+    return false;
+
+  if (!FBB) FBB = findFalseBlock(MBB, TBB);
+  if (!FBB)
+    // Malformed bcc? True and false blocks are the same?
+    return false;
+
+  // Restrict the optimization to cases where MBB is the only predecessor,
+  // it is an obvious win.
+  if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
+    return false;
+
+  // Find a suitable position to hoist the common instructions to. Also figure
+  // out which registers are used or defined by instructions from the insertion
+  // point to the end of the block.
+  SmallSet<unsigned, 4> Uses, Defs;
+  MachineBasicBlock::iterator Loc =
+    findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
+  if (Loc == MBB->end())
+    return false;
+
+  SmallSet<unsigned, 4> LocalDefs;
+  unsigned NumDups = 0;
+  MachineBasicBlock::iterator TIB = TBB->begin();
+  MachineBasicBlock::iterator FIB = FBB->begin();
+  MachineBasicBlock::iterator TIE = TBB->end();
+  MachineBasicBlock::iterator FIE = FBB->end();
+  while (TIB != TIE && FIB != FIE) {
+    // Skip dbg_value instructions. These do not count.
+    if (TIB->isDebugValue()) {
+      while (TIB != TIE && TIB->isDebugValue())
+        ++TIB;
+      if (TIB == TIE)
+        break;
+    }
+    if (FIB->isDebugValue()) {
+      while (FIB != FIE && FIB->isDebugValue())
+        ++FIB;
+      if (FIB == FIE)
+        break;
+    }
+    if (!TIB->isIdenticalTo(FIB))
+      break;
+
+    if (TII->isPredicated(TIB))
+      // Hard to reason about register liveness with predicated instruction.
+      break;
+
+    bool IsSafe = true;
+    for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
+      const MachineOperand &MO = TIB->getOperand(i);
+      if (!MO.isReg())
+        continue;
+      unsigned Reg = MO.getReg();
+      if (!Reg)
+        continue;
+      if (MO.isDef()) {
+        if (Uses.count(Reg)) {
+          // Avoid clobbering a register that's used by the instruction at
+          // the point of insertion.
+          IsSafe = false;
+          break;
+        }
+        if (!MO.isDead() && Defs.count(Reg)) {
+          // Don't hoist the instruction if the def would be clobber by the
+          // instruction at the point insertion. FIXME: This is overly
+          // conservative. It should be possible to hoist the instructions
+          // in BB2 in the following example:
+          // BB1:
+          // r1, eflag = op1 r2, r3
+          // brcc eflag
+          //
+          // BB2:
+          // r1 = op2, ...
+          //    = op3, r1<kill>
+          IsSafe = false;
+          break;
+        }
+        LocalDefs.insert(Reg);
+        for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
+          LocalDefs.insert(*SR);
+      } else if (!LocalDefs.count(Reg)) {
+        if (Defs.count(Reg)) {
+          // Use is defined by the instruction at the point of insertion.
+          IsSafe = false;
+          break;
+        }
+      }
+    }
+    if (!IsSafe)
+      break;
+
+    bool DontMoveAcrossStore = true;
+    if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore))
+      break;
+
+    ++NumDups;
+    ++TIB;
+    ++FIB;
+  }
+
+  if (!NumDups)
+    return false;
+
+  MBB->splice(Loc, TBB, TBB->begin(), TIB);
+  FBB->erase(FBB->begin(), FIB);
+  ++NumHoist;
+  return true;
+}
--- a/llvm/lib/CodeGen/BranchFolding.h
+++ b/llvm/lib/CodeGen/BranchFolding.h
@ -19,11 +19,10 @@ namespace llvm {
  class RegScavenger;
  class TargetInstrInfo;
  class TargetRegisterInfo;
-  template<typename T> class SmallVectorImpl;

  class BranchFolder {
  public:
-    explicit BranchFolder(bool defaultEnableTailMerge);
+    explicit BranchFolder(bool defaultEnableTailMerge, bool CommonHoist);

    bool OptimizeFunction(MachineFunction &MF,
                          const TargetInstrInfo *tii,
@ -85,6 +84,7 @@ namespace llvm {
    std::vector<SameTailElt> SameTails;

    bool EnableTailMerge;
+    bool EnableHoistCommonCode;
    const TargetInstrInfo *TII;
    const TargetRegisterInfo *TRI;
    MachineModuleInfo *MMI;
@ -110,6 +110,9 @@ namespace llvm {
    bool OptimizeBlock(MachineBasicBlock *MBB);
    void RemoveDeadBlock(MachineBasicBlock *MBB);
    bool OptimizeImpDefsBlock(MachineBasicBlock *MBB);
+
+    bool HoistCommonCode(MachineFunction &MF);
+    bool HoistCommonCodeInSuccs(MachineBasicBlock *MBB);
  };
 }

--- a/llvm/lib/CodeGen/IfConversion.cpp
+++ b/llvm/lib/CodeGen/IfConversion.cpp
@ -265,7 +265,7 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) {
  if (!TII) return false;

  // Tail merge tend to expose more if-conversion opportunities.
-  BranchFolder BF(true);
+  BranchFolder BF(true, false);
  bool BFChange = BF.OptimizeFunction(MF, TII,
                                   MF.getTarget().getRegisterInfo(),
                                   getAnalysisIfAvailable<MachineModuleInfo>());
@ -399,7 +399,7 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) {
  BBAnalysis.clear();

  if (MadeChange && IfCvtBranchFold) {
-    BranchFolder BF(false);
+    BranchFolder BF(false, false);
    BF.OptimizeFunction(MF, TII,
                        MF.getTarget().getRegisterInfo(),
                        getAnalysisIfAvailable<MachineModuleInfo>());
--- a/llvm/test/CodeGen/X86/hoist-common.ll
+++ b/llvm/test/CodeGen/X86/hoist-common.ll
@ -0,0 +1,28 @@
+; RUN: llc < %s -march=x86-64  | FileCheck %s
+
+; Common "xorb al, al" instruction in the two successor blocks should be
+; moved to the entry block above the test + je.
+
+; rdar://9145558
+
+define zeroext i1 @t(i32 %c) nounwind ssp {
+entry:
+; CHECK: t:
+; CHECK: xorb %al, %al
+; CHECK: test
+; CHECK: je
+  %tobool = icmp eq i32 %c, 0
+  br i1 %tobool, label %return, label %if.then
+
+if.then:
+; CHECK: callq
+  %call = tail call zeroext i1 (...)* @foo() nounwind
+  br label %return
+
+return:
+; CHECK: ret
+  %retval.0 = phi i1 [ %call, %if.then ], [ false, %entry ]
+  ret i1 %retval.0
+}
+
+declare zeroext i1 @foo(...)