//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "branchfolding"
#include "llvm/CodeGen/Passes.h"
-#include "llvm/CodeGen/MachineDebugInfo.h"
+#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
+#include "llvm/CodeGen/RegisterScavenging.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/MRegisterInfo.h"
#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
+#include <algorithm>
using namespace llvm;
-static Statistic<> NumDeadBlocks("branchfold", "Number of dead blocks removed");
-static Statistic<> NumBranchOpts("branchfold", "Number of branches optimized");
-static Statistic<> NumTailMerge ("branchfold", "Number of block tails merged");
-static cl::opt<bool> EnableTailMerge("enable-tail-merge", cl::init(false));
+STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
+STATISTIC(NumBranchOpts, "Number of branches optimized");
+STATISTIC(NumTailMerge , "Number of block tails merged");
+static cl::opt<bool> EnableTailMerge("enable-tail-merge", cl::Hidden);
namespace {
struct BranchFolder : public MachineFunctionPass {
virtual bool runOnMachineFunction(MachineFunction &MF);
virtual const char *getPassName() const { return "Control Flow Optimizer"; }
const TargetInstrInfo *TII;
- MachineDebugInfo *MDI;
+ MachineModuleInfo *MMI;
bool MadeChange;
private:
// Tail Merging.
bool TailMergeBlocks(MachineFunction &MF);
void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
MachineBasicBlock *NewDest);
+ MachineBasicBlock *SplitMBBAt(MachineBasicBlock &CurMBB,
+ MachineBasicBlock::iterator BBI1);
+ const MRegisterInfo *RegInfo;
+ RegScavenger *RS;
// Branch optzn.
bool OptimizeBranches(MachineFunction &MF);
- void OptimizeBlock(MachineFunction::iterator MBB);
+ void OptimizeBlock(MachineBasicBlock *MBB);
void RemoveDeadBlock(MachineBasicBlock *MBB);
+
+ bool CanFallThrough(MachineBasicBlock *CurBB);
+ bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable,
+ MachineBasicBlock *TBB, MachineBasicBlock *FBB,
+ const std::vector<MachineOperand> &Cond);
};
}
/// function, updating the CFG.
void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
assert(MBB->pred_empty() && "MBB must be dead!");
+ DOUT << "\nRemoving MBB: " << *MBB;
MachineFunction *MF = MBB->getParent();
// drop all successors.
while (!MBB->succ_empty())
MBB->removeSuccessor(MBB->succ_end()-1);
- // If there is DWARF info to active, check to see if there are any DWARF_LABEL
- // records in the basic block. If so, unregister them from MachineDebugInfo.
- if (MDI && !MBB->empty()) {
- unsigned DWARF_LABELOpc = TII->getDWARF_LABELOpcode();
- assert(DWARF_LABELOpc &&
- "Target supports dwarf but didn't implement getDWARF_LABELOpcode!");
-
+ // If there is DWARF info to active, check to see if there are any LABEL
+ // records in the basic block. If so, unregister them from MachineModuleInfo.
+ if (MMI && !MBB->empty()) {
for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
I != E; ++I) {
- if ((unsigned)I->getOpcode() == DWARF_LABELOpc) {
+ if ((unsigned)I->getOpcode() == TargetInstrInfo::LABEL) {
// The label ID # is always operand #0, an immediate.
- MDI->RemoveLabelInfo(I->getOperand(0).getImm());
+ MMI->InvalidateLabel(I->getOperand(0).getImm());
}
}
}
TII = MF.getTarget().getInstrInfo();
if (!TII) return false;
- MDI = getAnalysisToUpdate<MachineDebugInfo>();
+ RegInfo = MF.getTarget().getRegisterInfo();
+ RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL;
+
+ MMI = getAnalysisToUpdate<MachineModuleInfo>();
bool EverMadeChange = false;
bool MadeChangeThisIteration = true;
EverMadeChange |= MadeChangeThisIteration;
}
+ // See if any jump tables have become mergable or dead as the code generator
+ // did its thing.
+ MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
+ const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
+ if (!JTs.empty()) {
+ // Figure out how these jump tables should be merged.
+ std::vector<unsigned> JTMapping;
+ JTMapping.reserve(JTs.size());
+
+ // We always keep the 0th jump table.
+ JTMapping.push_back(0);
+
+ // Scan the jump tables, seeing if there are any duplicates. Note that this
+ // is N^2, which should be fixed someday.
+ for (unsigned i = 1, e = JTs.size(); i != e; ++i)
+ JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
+
+ // If a jump table was merge with another one, walk the function rewriting
+ // references to jump tables to reference the new JT ID's. Keep track of
+ // whether we see a jump table idx, if not, we can delete the JT.
+ std::vector<bool> JTIsLive;
+ JTIsLive.resize(JTs.size());
+ for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
+ BB != E; ++BB) {
+ for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
+ I != E; ++I)
+ for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
+ MachineOperand &Op = I->getOperand(op);
+ if (!Op.isJumpTableIndex()) continue;
+ unsigned NewIdx = JTMapping[Op.getJumpTableIndex()];
+ Op.setJumpTableIndex(NewIdx);
+
+ // Remember that this JT is live.
+ JTIsLive[NewIdx] = true;
+ }
+ }
+
+ // Finally, remove dead jump tables. This happens either because the
+ // indirect jump was unreachable (and thus deleted) or because the jump
+ // table was merged with some other one.
+ for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
+ if (!JTIsLive[i]) {
+ JTI->RemoveJumpTable(i);
+ EverMadeChange = true;
+ }
+ }
+
+ delete RS;
return EverMadeChange;
}
++NumTailMerge;
}
+/// SplitMBBAt - Given a machine basic block and an iterator into it, split the
+/// MBB so that the part before the iterator falls into the part starting at the
+/// iterator. This returns the new MBB.
+MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
+ MachineBasicBlock::iterator BBI1) {
+ // Create the fall-through block.
+ MachineFunction::iterator MBBI = &CurMBB;
+ MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock());
+ CurMBB.getParent()->getBasicBlockList().insert(++MBBI, NewMBB);
+
+ // Move all the successors of this block to the specified block.
+ while (!CurMBB.succ_empty()) {
+ MachineBasicBlock *S = *(CurMBB.succ_end()-1);
+ NewMBB->addSuccessor(S);
+ CurMBB.removeSuccessor(S);
+ }
+
+ // Add an edge from CurMBB to NewMBB for the fall-through.
+ CurMBB.addSuccessor(NewMBB);
+
+ // Splice the code over.
+ NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
+
+ // For targets that use the register scavenger, we must maintain LiveIns.
+ if (RS) {
+ RS->enterBasicBlock(&CurMBB);
+ if (!CurMBB.empty())
+ RS->forward(prior(CurMBB.end()));
+ BitVector RegsLiveAtExit(RegInfo->getNumRegs());
+ RS->getRegsUsed(RegsLiveAtExit, false);
+ for (unsigned int i=0, e=RegInfo->getNumRegs(); i!=e; i++)
+ if (RegsLiveAtExit[i])
+ NewMBB->addLiveIn(i);
+ }
+
+ return NewMBB;
+}
+
+/// EstimateRuntime - Make a rough estimate for how long it will take to run
+/// the specified code.
+static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
+ MachineBasicBlock::iterator E,
+ const TargetInstrInfo *TII) {
+ unsigned Time = 0;
+ for (; I != E; ++I) {
+ const TargetInstrDescriptor &TID = TII->get(I->getOpcode());
+ if (TID.Flags & M_CALL_FLAG)
+ Time += 10;
+ else if (TID.Flags & (M_LOAD_FLAG|M_STORE_FLAG))
+ Time += 2;
+ else
+ ++Time;
+ }
+ return Time;
+}
+
+/// ShouldSplitFirstBlock - We need to either split MBB1 at MBB1I or MBB2 at
+/// MBB2I and then insert an unconditional branch in the other block. Determine
+/// which is the best to split
+static bool ShouldSplitFirstBlock(MachineBasicBlock *MBB1,
+ MachineBasicBlock::iterator MBB1I,
+ MachineBasicBlock *MBB2,
+ MachineBasicBlock::iterator MBB2I,
+ const TargetInstrInfo *TII) {
+ // TODO: if we had some notion of which block was hotter, we could split
+ // the hot block, so it is the fall-through. Since we don't have profile info
+ // make a decision based on which will hurt most to split.
+ unsigned MBB1Time = EstimateRuntime(MBB1->begin(), MBB1I, TII);
+ unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I, TII);
+
+ // If the MBB1 prefix takes "less time" to run than the MBB2 prefix, split the
+ // MBB1 block so it falls through. This will penalize the MBB2 path, but will
+ // have a lower overall impact on the program execution.
+ return MBB1Time < MBB2Time;
+}
+
bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
MadeChange = false;
- if (!EnableTailMerge)
- return false;
+ if (!EnableTailMerge) return false;
// Find blocks with no successors.
std::vector<std::pair<unsigned,MachineBasicBlock*> > MergePotentials;
// Otherwise, move the matching block to the right position.
std::swap(MergePotentials[FoundMatch], *(MergePotentials.end()-2));
}
-
- // If either block is the entire common tail, make the longer one branch to
- // the shorter one.
+
MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second;
- if (CurMBB->begin() == BBI1) {
- // Hack the end off MBB2, making it jump to CurMBB instead.
- ReplaceTailWithBranchTo(BBI2, CurMBB);
- // This modifies MBB2, so remove it from the worklist.
- MergePotentials.erase(MergePotentials.end()-2);
- MadeChange = true;
- continue;
- } else if (MBB2->begin() == BBI2) {
+
+ // If neither block is the entire common tail, split the tail of one block
+ // to make it redundant with the other tail.
+ if (CurMBB->begin() != BBI1 && MBB2->begin() != BBI2) {
+ if (0) { // Enable this to disable partial tail merges.
+ MergePotentials.pop_back();
+ continue;
+ }
+
+ // Decide whether we want to split CurMBB or MBB2.
+ if (ShouldSplitFirstBlock(CurMBB, BBI1, MBB2, BBI2, TII)) {
+ CurMBB = SplitMBBAt(*CurMBB, BBI1);
+ BBI1 = CurMBB->begin();
+ MergePotentials.back().second = CurMBB;
+ } else {
+ MBB2 = SplitMBBAt(*MBB2, BBI2);
+ BBI2 = MBB2->begin();
+ (MergePotentials.end()-2)->second = MBB2;
+ }
+ }
+
+ if (MBB2->begin() == BBI2) {
// Hack the end off CurMBB, making it jump to MBBI@ instead.
ReplaceTailWithBranchTo(BBI1, MBB2);
// This modifies CurMBB, so remove it from the worklist.
MergePotentials.pop_back();
- MadeChange = true;
- continue;
+ } else {
+ assert(CurMBB->begin() == BBI1 && "Didn't split block correctly?");
+ // Hack the end off MBB2, making it jump to CurMBB instead.
+ ReplaceTailWithBranchTo(BBI2, CurMBB);
+ // This modifies MBB2, so remove it from the worklist.
+ MergePotentials.erase(MergePotentials.end()-2);
}
-
- MergePotentials.pop_back();
+ MadeChange = true;
}
return MadeChange;
bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
MadeChange = false;
+ // Make sure blocks are numbered in order
+ MF.RenumberBlocks();
+
for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
MachineBasicBlock *MBB = I++;
OptimizeBlock(MBB);
} else if (*SI == DestB) {
DestB = 0;
++SI;
+ } else if ((*SI)->isLandingPad()) {
+ ++SI;
} else {
// Otherwise, this is a superfluous edge, remove it.
MBB.removeSuccessor(SI);
}
}
+/// CanFallThrough - Return true if the specified block (with the specified
+/// branch condition) can implicitly transfer control to the block after it by
+/// falling off the end of it. This should return false if it can reach the
+/// block after it, but it uses an explicit branch to do so (e.g. a table jump).
+///
+/// True is a conservative answer.
+///
+bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB,
+ bool BranchUnAnalyzable,
+ MachineBasicBlock *TBB, MachineBasicBlock *FBB,
+ const std::vector<MachineOperand> &Cond) {
+ MachineFunction::iterator Fallthrough = CurBB;
+ ++Fallthrough;
+ // If FallthroughBlock is off the end of the function, it can't fall through.
+ if (Fallthrough == CurBB->getParent()->end())
+ return false;
+
+ // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible.
+ if (!CurBB->isSuccessor(Fallthrough))
+ return false;
+
+ // If we couldn't analyze the branch, assume it could fall through.
+ if (BranchUnAnalyzable) return true;
+
+ // If there is no branch, control always falls through.
+ if (TBB == 0) return true;
+
+ // If there is some explicit branch to the fallthrough block, it can obviously
+ // reach, even though the branch should get folded to fall through implicitly.
+ if (MachineFunction::iterator(TBB) == Fallthrough ||
+ MachineFunction::iterator(FBB) == Fallthrough)
+ return true;
+
+ // If it's an unconditional branch to some block not the fall through, it
+ // doesn't fall through.
+ if (Cond.empty()) return false;
+
+ // Otherwise, if it is conditional and has no explicit false block, it falls
+ // through.
+ return FBB == 0;
+}
+
+/// CanFallThrough - Return true if the specified can implicitly transfer
+/// control to the block after it by falling off the end of it. This should
+/// return false if it can reach the block after it, but it uses an explicit
+/// branch to do so (e.g. a table jump).
+///
+/// True is a conservative answer.
+///
+bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) {
+ MachineBasicBlock *TBB = 0, *FBB = 0;
+ std::vector<MachineOperand> Cond;
+ bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond);
+ return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
+}
+
+/// IsBetterFallthrough - Return true if it would be clearly better to
+/// fall-through to MBB1 than to fall through into MBB2. This has to return
+/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
+/// result in infinite loops.
+static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
+ MachineBasicBlock *MBB2,
+ const TargetInstrInfo &TII) {
+ // Right now, we use a simple heuristic. If MBB2 ends with a call, and
+ // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
+ // optimize branches that branch to either a return block or an assert block
+ // into a fallthrough to the return.
+ if (MBB1->empty() || MBB2->empty()) return false;
+
+ MachineInstr *MBB1I = --MBB1->end();
+ MachineInstr *MBB2I = --MBB2->end();
+ return TII.isCall(MBB2I->getOpcode()) && !TII.isCall(MBB1I->getOpcode());
+}
+
/// OptimizeBlock - Analyze and optimize control flow related to the specified
/// block. This is never called on the entry block.
-void BranchFolder::OptimizeBlock(MachineFunction::iterator MBB) {
+void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
+ MachineFunction::iterator FallThrough = MBB;
+ ++FallThrough;
+
// If this block is empty, make everyone use its fall-through, not the block
// explicitly.
if (MBB->empty()) {
// Dead block? Leave for cleanup later.
if (MBB->pred_empty()) return;
- MachineFunction::iterator FallThrough = next(MBB);
-
if (FallThrough == MBB->getParent()->end()) {
// TODO: Simplify preds to not branch here if possible!
} else {
// If MBB was the target of a jump table, update jump tables to go to the
// fallthrough instead.
- MBB->getParent()->getJumpTableInfo()->ReplaceMBBInJumpTables(MBB,
- FallThrough);
+ MBB->getParent()->getJumpTableInfo()->
+ ReplaceMBBInJumpTables(MBB, FallThrough);
MadeChange = true;
}
return;
// Check to see if we can simplify the terminator of the block before this
// one.
- MachineBasicBlock &PrevBB = *prior(MBB);
+ MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
std::vector<MachineOperand> PriorCond;
- bool PriorUnAnalyzable = false;
- PriorUnAnalyzable = TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
+ bool PriorUnAnalyzable =
+ TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
if (!PriorUnAnalyzable) {
// If the CFG for the prior block has extra edges, remove them.
MadeChange |= CorrectExtraCFGEdges(PrevBB, PriorTBB, PriorFBB,
if (PriorTBB && PriorTBB == PriorFBB) {
TII->RemoveBranch(PrevBB);
PriorCond.clear();
- if (PriorTBB != &*MBB)
+ if (PriorTBB != MBB)
TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
MadeChange = true;
++NumBranchOpts;
// If the previous branch *only* branches to *this* block (conditional or
// not) remove the branch.
- if (PriorTBB == &*MBB && PriorFBB == 0) {
+ if (PriorTBB == MBB && PriorFBB == 0) {
TII->RemoveBranch(PrevBB);
MadeChange = true;
++NumBranchOpts;
// If the prior block branches somewhere else on the condition and here if
// the condition is false, remove the uncond second branch.
- if (PriorFBB == &*MBB) {
+ if (PriorFBB == MBB) {
TII->RemoveBranch(PrevBB);
TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
MadeChange = true;
// If the prior block branches here on true and somewhere else on false, and
// if the branch condition is reversible, reverse the branch to create a
// fall-through.
- if (PriorTBB == &*MBB) {
+ if (PriorTBB == MBB) {
std::vector<MachineOperand> NewPriorCond(PriorCond);
if (!TII->ReverseBranchCondition(NewPriorCond)) {
TII->RemoveBranch(PrevBB);
return OptimizeBlock(MBB);
}
}
+
+ // If this block doesn't fall through (e.g. it ends with an uncond branch or
+ // has no successors) and if the pred falls through into this block, and if
+ // it would otherwise fall through into the block after this, move this
+ // block to the end of the function.
+ //
+ // We consider it more likely that execution will stay in the function (e.g.
+ // due to loops) than it is to exit it. This asserts in loops etc, moving
+ // the assert condition out of the loop body.
+ if (!PriorCond.empty() && PriorFBB == 0 &&
+ MachineFunction::iterator(PriorTBB) == FallThrough &&
+ !CanFallThrough(MBB)) {
+ bool DoTransform = true;
+
+ // We have to be careful that the succs of PredBB aren't both no-successor
+ // blocks. If neither have successors and if PredBB is the second from
+ // last block in the function, we'd just keep swapping the two blocks for
+ // last. Only do the swap if one is clearly better to fall through than
+ // the other.
+ if (FallThrough == --MBB->getParent()->end() &&
+ !IsBetterFallthrough(PriorTBB, MBB, *TII))
+ DoTransform = false;
+
+ // We don't want to do this transformation if we have control flow like:
+ // br cond BB2
+ // BB1:
+ // ..
+ // jmp BBX
+ // BB2:
+ // ..
+ // ret
+ //
+ // In this case, we could actually be moving the return block *into* a
+ // loop!
+ if (DoTransform && !MBB->succ_empty() &&
+ (!CanFallThrough(PriorTBB) || PriorTBB->empty()))
+ DoTransform = false;
+
+
+ if (DoTransform) {
+ // Reverse the branch so we will fall through on the previous true cond.
+ std::vector<MachineOperand> NewPriorCond(PriorCond);
+ if (!TII->ReverseBranchCondition(NewPriorCond)) {
+ DOUT << "\nMoving MBB: " << *MBB;
+ DOUT << "To make fallthrough to: " << *PriorTBB << "\n";
+
+ TII->RemoveBranch(PrevBB);
+ TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
+
+ // Move this block to the end of the function.
+ MBB->moveAfter(--MBB->getParent()->end());
+ MadeChange = true;
+ ++NumBranchOpts;
+ return;
+ }
+ }
+ }
}
// Analyze the branch in the current block.
MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
std::vector<MachineOperand> CurCond;
- if (!TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond)) {
+ bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond);
+ if (!CurUnAnalyzable) {
// If the CFG for the prior block has extra edges, remove them.
MadeChange |= CorrectExtraCFGEdges(*MBB, CurTBB, CurFBB,
- !CurCond.empty(), next(MBB));
-
+ !CurCond.empty(),
+ ++MachineFunction::iterator(MBB));
+
+ // If this is a two-way branch, and the FBB branches to this block, reverse
+ // the condition so the single-basic-block loop is faster. Instead of:
+ // Loop: xxx; jcc Out; jmp Loop
+ // we want:
+ // Loop: xxx; jncc Loop; jmp Out
+ if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
+ std::vector<MachineOperand> NewCond(CurCond);
+ if (!TII->ReverseBranchCondition(NewCond)) {
+ TII->RemoveBranch(*MBB);
+ TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
+ MadeChange = true;
+ ++NumBranchOpts;
+ return OptimizeBlock(MBB);
+ }
+ }
+
+
// If this branch is the only thing in its block, see if we can forward
// other blocks across it.
if (CurTBB && CurCond.empty() && CurFBB == 0 &&
- TII->isBranch(MBB->begin()->getOpcode())) {
+ TII->isBranch(MBB->begin()->getOpcode()) && CurTBB != MBB) {
// This block may contain just an unconditional branch. Because there can
// be 'non-branch terminators' in the block, try removing the branch and
// then seeing if the block is empty.
// completely eliminate the block is when the block before this one
// falls through into MBB and we can't understand the prior block's branch
// condition.
- if (MBB->empty() && (!PriorUnAnalyzable || !PrevBB.isSuccessor(MBB))) {
- // If the prior block falls through into us, turn it into an
- // explicit branch to us to make updates simpler.
- if (PrevBB.isSuccessor(MBB) && PriorTBB != &*MBB && PriorFBB != &*MBB) {
- if (PriorTBB == 0) {
- assert(PriorCond.empty() && PriorFBB == 0 && "Bad branch analysis");
- PriorTBB = MBB;
- } else {
- assert(PriorFBB == 0 && "Machine CFG out of date!");
- PriorFBB = MBB;
+ if (MBB->empty()) {
+ bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
+ if (PredHasNoFallThrough || !PriorUnAnalyzable ||
+ !PrevBB.isSuccessor(MBB)) {
+ // If the prior block falls through into us, turn it into an
+ // explicit branch to us to make updates simpler.
+ if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
+ PriorTBB != MBB && PriorFBB != MBB) {
+ if (PriorTBB == 0) {
+ assert(PriorCond.empty() && PriorFBB == 0 &&
+ "Bad branch analysis");
+ PriorTBB = MBB;
+ } else {
+ assert(PriorFBB == 0 && "Machine CFG out of date!");
+ PriorFBB = MBB;
+ }
+ TII->RemoveBranch(PrevBB);
+ TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
}
- TII->RemoveBranch(PrevBB);
- TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
- }
- // Iterate through all the predecessors, revectoring each in-turn.
- while (!MBB->pred_empty())
- ReplaceUsesOfBlockWith(*(MBB->pred_end()-1), MBB, CurTBB, TII);
+ // Iterate through all the predecessors, revectoring each in-turn.
+ MachineBasicBlock::pred_iterator PI = MBB->pred_begin();
+ bool DidChange = false;
+ bool HasBranchToSelf = false;
+ while (PI != MBB->pred_end()) {
+ if (*PI == MBB) {
+ // If this block has an uncond branch to itself, leave it.
+ ++PI;
+ HasBranchToSelf = true;
+ } else {
+ DidChange = true;
+ ReplaceUsesOfBlockWith(*PI, MBB, CurTBB, TII);
+ }
+ }
- // Change any jumptables to go to the new MBB.
- MBB->getParent()->getJumpTableInfo()->ReplaceMBBInJumpTables(MBB,
- CurTBB);
- ++NumBranchOpts;
- MadeChange = true;
- return;
+ // Change any jumptables to go to the new MBB.
+ MBB->getParent()->getJumpTableInfo()->
+ ReplaceMBBInJumpTables(MBB, CurTBB);
+ if (DidChange) {
+ ++NumBranchOpts;
+ MadeChange = true;
+ if (!HasBranchToSelf) return;
+ }
+ }
}
// Add the branch back if the block is more than just an uncond branch.
TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
}
}
+
+ // If the prior block doesn't fall through into this block, and if this
+ // block doesn't fall through into some other block, see if we can find a
+ // place to move this block where a fall-through will happen.
+ if (!CanFallThrough(&PrevBB, PriorUnAnalyzable,
+ PriorTBB, PriorFBB, PriorCond)) {
+ // Now we know that there was no fall-through into this block, check to
+ // see if it has a fall-through into its successor.
+ bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB,
+ CurCond);
+
+ if (!MBB->isLandingPad()) {
+ // Check all the predecessors of this block. If one of them has no fall
+ // throughs, move this block right after it.
+ for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
+ E = MBB->pred_end(); PI != E; ++PI) {
+ // Analyze the branch at the end of the pred.
+ MachineBasicBlock *PredBB = *PI;
+ MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
+ if (PredBB != MBB && !CanFallThrough(PredBB)
+ && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
+ // If the current block doesn't fall through, just move it.
+ // If the current block can fall through and does not end with a
+ // conditional branch, we need to append an unconditional jump to
+ // the (current) next block. To avoid a possible compile-time
+ // infinite loop, move blocks only backward in this case.
+ if (CurFallsThru) {
+ MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
+ CurCond.clear();
+ TII->InsertBranch(*MBB, NextBB, 0, CurCond);
+ }
+ MBB->moveAfter(PredBB);
+ MadeChange = true;
+ return OptimizeBlock(MBB);
+ }
+ }
+ }
+
+ if (!CurFallsThru) {
+ // Check all successors to see if we can move this block before it.
+ for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
+ E = MBB->succ_end(); SI != E; ++SI) {
+ // Analyze the branch at the end of the block before the succ.
+ MachineBasicBlock *SuccBB = *SI;
+ MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
+ std::vector<MachineOperand> SuccPrevCond;
+ if (SuccBB != MBB && !CanFallThrough(SuccPrev)) {
+ MBB->moveBefore(SuccBB);
+ MadeChange = true;
+ return OptimizeBlock(MBB);
+ }
+ }
+
+ // Okay, there is no really great place to put this block. If, however,
+ // the block before this one would be a fall-through if this block were
+ // removed, move this block to the end of the function.
+ if (FallThrough != MBB->getParent()->end() &&
+ PrevBB.isSuccessor(FallThrough)) {
+ MBB->moveAfter(--MBB->getParent()->end());
+ MadeChange = true;
+ return;
+ }
+ }
+ }
}