#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
-#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/ADT/Statistic.h"
using namespace llvm;
-static cl::opt<bool>
-OptLoopBBPlacement("opt-loop-bb-placement",
- cl::init(false), cl::Hidden,
- cl::desc("Optimize block placements in loops"));
-
-STATISTIC(NumHeaderAligned, "Number of loop header aligned");
+STATISTIC(NumLoopsAligned, "Number of loops aligned");
STATISTIC(NumIntraElim, "Number of intra loop branches eliminated");
STATISTIC(NumIntraMoved, "Number of intra loop branches moved");
const TargetInstrInfo *TII;
const TargetLowering *TLI;
- /// ChangedMBBs - BBs which are modified by OptimizeIntraLoopEdges.
- SmallPtrSet<MachineBasicBlock*, 8> ChangedMBBs;
-
- /// UncondJmpMBBs - A list of BBs which are in loops and end with
- /// unconditional branches.
- SmallVector<std::pair<MachineBasicBlock*,MachineBasicBlock*>, 4>
- UncondJmpMBBs;
-
- /// LoopHeaders - A list of BBs which are loop headers.
- SmallVector<MachineBasicBlock*, 4> LoopHeaders;
-
public:
static char ID;
- CodePlacementOpt() : MachineFunctionPass(&ID) {}
+ CodePlacementOpt() : MachineFunctionPass(ID) {}
virtual bool runOnMachineFunction(MachineFunction &MF);
virtual const char *getPassName() const {
}
private:
- bool OptimizeIntraLoopEdges();
+ bool HasFallthrough(MachineBasicBlock *MBB);
+ bool HasAnalyzableTerminator(MachineBasicBlock *MBB);
+ void Splice(MachineFunction &MF,
+ MachineFunction::iterator InsertPt,
+ MachineFunction::iterator Begin,
+ MachineFunction::iterator End);
+ bool EliminateUnconditionalJumpsToTop(MachineFunction &MF,
+ MachineLoop *L);
+ bool MoveDiscontiguousLoopBlocks(MachineFunction &MF,
+ MachineLoop *L);
+ bool OptimizeIntraLoopEdgesInLoopNest(MachineFunction &MF, MachineLoop *L);
+ bool OptimizeIntraLoopEdges(MachineFunction &MF);
bool AlignLoops(MachineFunction &MF);
+ bool AlignLoop(MachineFunction &MF, MachineLoop *L, unsigned Align);
};
char CodePlacementOpt::ID = 0;
return new CodePlacementOpt();
}
-/// OptimizeBackEdges - Place loop back edges to move unconditional branches
-/// out of the loop.
-///
-/// A:
-/// ...
-/// <fallthrough to B>
-///
-/// B: --> loop header
-/// ...
-/// jcc <cond> C, [exit]
+/// HasFallthrough - Test whether the given branch has a fallthrough, either as
+/// a plain fallthrough or as a fallthrough case of a conditional branch.
///
-/// C:
-/// ...
-/// jmp B
-///
-/// ==>
-///
-/// A:
-/// ...
-/// jmp B
+bool CodePlacementOpt::HasFallthrough(MachineBasicBlock *MBB) {
+ MachineBasicBlock *TBB = 0, *FBB = 0;
+ SmallVector<MachineOperand, 4> Cond;
+ if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond))
+ return false;
+ // This conditional branch has no fallthrough.
+ if (FBB)
+ return false;
+ // An unconditional branch has no fallthrough.
+ if (Cond.empty() && TBB)
+ return false;
+ // It has a fallthrough.
+ return true;
+}
+
+/// HasAnalyzableTerminator - Test whether AnalyzeBranch will succeed on MBB.
+/// This is called before major changes are begun to test whether it will be
+/// possible to complete the changes.
///
-/// C: --> new loop header
-/// ...
-/// <fallthough to B>
-///
-/// B:
-/// ...
-/// jcc <cond> C, [exit]
+/// Target-specific code is hereby encouraged to make AnalyzeBranch succeed
+/// whenever possible.
///
-bool CodePlacementOpt::OptimizeIntraLoopEdges() {
- if (!OptLoopBBPlacement)
+bool CodePlacementOpt::HasAnalyzableTerminator(MachineBasicBlock *MBB) {
+ // Conservatively ignore EH landing pads.
+ if (MBB->isLandingPad()) return false;
+
+ // Aggressively handle return blocks and similar constructs.
+ if (MBB->succ_empty()) return true;
+
+ // Ask the target's AnalyzeBranch if it can handle this block.
+ MachineBasicBlock *TBB = 0, *FBB = 0;
+ SmallVector<MachineOperand, 4> Cond;
+ // Make sure the terminator is understood.
+ if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond))
+ return false;
+ // Ignore blocks which look like they might have EH-related control flow.
+ // AnalyzeBranch thinks it knows how to analyze such things, but it doesn't
+ // recognize the possibility of a control transfer through an unwind.
+ // Such blocks contain EH_LABEL instructions, however they may be in the
+ // middle of the block. Instead of searching for them, just check to see
+ // if the CFG disagrees with AnalyzeBranch.
+ if (1u + !Cond.empty() != MBB->succ_size())
return false;
+ // Make sure we have the option of reversing the condition.
+ if (!Cond.empty() && TII->ReverseBranchCondition(Cond))
+ return false;
+ return true;
+}
+/// Splice - Move the sequence of instructions [Begin,End) to just before
+/// InsertPt. Update branch instructions as needed to account for broken
+/// fallthrough edges and to take advantage of newly exposed fallthrough
+/// opportunities.
+///
+void CodePlacementOpt::Splice(MachineFunction &MF,
+ MachineFunction::iterator InsertPt,
+ MachineFunction::iterator Begin,
+ MachineFunction::iterator End) {
+ assert(Begin != MF.begin() && End != MF.begin() && InsertPt != MF.begin() &&
+ "Splice can't change the entry block!");
+ MachineFunction::iterator OldBeginPrior = prior(Begin);
+ MachineFunction::iterator OldEndPrior = prior(End);
+
+ MF.splice(InsertPt, Begin, End);
+
+ prior(Begin)->updateTerminator();
+ OldBeginPrior->updateTerminator();
+ OldEndPrior->updateTerminator();
+}
+
+/// EliminateUnconditionalJumpsToTop - Move blocks which unconditionally jump
+/// to the loop top to the top of the loop so that they have a fall through.
+/// This can introduce a branch on entry to the loop, but it can eliminate a
+/// branch within the loop. See the @simple case in
+/// test/CodeGen/X86/loop_blocks.ll for an example of this.
+bool CodePlacementOpt::EliminateUnconditionalJumpsToTop(MachineFunction &MF,
+ MachineLoop *L) {
bool Changed = false;
- for (unsigned i = 0, e = UncondJmpMBBs.size(); i != e; ++i) {
- MachineBasicBlock *MBB = UncondJmpMBBs[i].first;
- MachineBasicBlock *SuccMBB = UncondJmpMBBs[i].second;
- MachineLoop *L = MLI->getLoopFor(MBB);
- assert(L && "BB is expected to be in a loop!");
-
- if (ChangedMBBs.count(MBB)) {
- // BB has been modified, re-analyze.
- MachineBasicBlock *TBB = 0, *FBB = 0;
- SmallVector<MachineOperand, 4> Cond;
- if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond) || !Cond.empty())
+ MachineBasicBlock *TopMBB = L->getTopBlock();
+
+ bool BotHasFallthrough = HasFallthrough(L->getBottomBlock());
+
+ if (TopMBB == MF.begin() ||
+ HasAnalyzableTerminator(prior(MachineFunction::iterator(TopMBB)))) {
+ new_top:
+ for (MachineBasicBlock::pred_iterator PI = TopMBB->pred_begin(),
+ PE = TopMBB->pred_end(); PI != PE; ++PI) {
+ MachineBasicBlock *Pred = *PI;
+ if (Pred == TopMBB) continue;
+ if (HasFallthrough(Pred)) continue;
+ if (!L->contains(Pred)) continue;
+
+ // Verify that we can analyze all the loop entry edges before beginning
+ // any changes which will require us to be able to analyze them.
+ if (Pred == MF.begin())
continue;
- if (MLI->getLoopFor(TBB) != L || TBB->isLandingPad())
+ if (!HasAnalyzableTerminator(Pred))
+ continue;
+ if (!HasAnalyzableTerminator(prior(MachineFunction::iterator(Pred))))
continue;
- SuccMBB = TBB;
- } else {
- assert(MLI->getLoopFor(SuccMBB) == L &&
- "Successor is not in the same loop!");
- }
- if (MBB->isLayoutSuccessor(SuccMBB)) {
- // Successor is right after MBB, just eliminate the unconditional jmp.
- // Can this happen?
- TII->RemoveBranch(*MBB);
- ChangedMBBs.insert(MBB);
- ++NumIntraElim;
- continue;
- }
+ // Move the block.
+ DEBUG(dbgs() << "CGP: Moving blocks starting at BB#" << Pred->getNumber()
+ << " to top of loop.\n");
+ Changed = true;
- // Now check if the predecessor is fallthrough from any BB. If there is,
- // that BB should be from outside the loop since edge will become a jmp.
- bool OkToMove = true;
- MachineBasicBlock *FtMBB = 0, *FtTBB = 0, *FtFBB = 0;
- SmallVector<MachineOperand, 4> FtCond;
- for (MachineBasicBlock::pred_iterator PI = SuccMBB->pred_begin(),
- PE = SuccMBB->pred_end(); PI != PE; ++PI) {
- MachineBasicBlock *PredMBB = *PI;
- if (PredMBB->isLayoutSuccessor(SuccMBB)) {
- if (TII->AnalyzeBranch(*PredMBB, FtTBB, FtFBB, FtCond)) {
- OkToMove = false;
+ // Move it and all the blocks that can reach it via fallthrough edges
+ // exclusively, to keep existing fallthrough edges intact.
+ MachineFunction::iterator Begin = Pred;
+ MachineFunction::iterator End = llvm::next(Begin);
+ while (Begin != MF.begin()) {
+ MachineFunction::iterator Prior = prior(Begin);
+ if (Prior == MF.begin())
break;
- }
- if (!FtTBB)
- FtTBB = SuccMBB;
- else if (!FtFBB) {
- assert(FtFBB != SuccMBB && "Unexpected control flow!");
- FtFBB = SuccMBB;
- }
-
- // A fallthrough.
- FtMBB = PredMBB;
- MachineLoop *PL = MLI->getLoopFor(PredMBB);
- if (PL && (PL == L || PL->getLoopDepth() >= L->getLoopDepth())) {
- OkToMove = false;
+ // Stop when a non-fallthrough edge is found.
+ if (!HasFallthrough(Prior))
+ break;
+ // Stop if a block which could fall-through out of the loop is found.
+ if (Prior->isSuccessor(End))
+ break;
+ // If we've reached the top, stop scanning.
+ if (Prior == MachineFunction::iterator(TopMBB)) {
+ // We know top currently has a fall through (because we just checked
+ // it) which would be lost if we do the transformation, so it isn't
+ // worthwhile to do the transformation unless it would expose a new
+ // fallthrough edge.
+ if (!Prior->isSuccessor(End))
+ goto next_pred;
+ // Otherwise we can stop scanning and procede to move the blocks.
break;
}
+ // If we hit a switch or something complicated, don't move anything
+ // for this predecessor.
+ if (!HasAnalyzableTerminator(prior(MachineFunction::iterator(Prior))))
+ break;
+ // Ok, the block prior to Begin will be moved along with the rest.
+ // Extend the range to include it.
+ Begin = Prior;
+ ++NumIntraMoved;
}
+
+ // Move the blocks.
+ Splice(MF, TopMBB, Begin, End);
+
+ // Update TopMBB.
+ TopMBB = L->getTopBlock();
+
+ // We have a new loop top. Iterate on it. We shouldn't have to do this
+ // too many times if BranchFolding has done a reasonable job.
+ goto new_top;
+ next_pred:;
}
+ }
- if (!OkToMove)
- continue;
-
- // Is it profitable? If SuccMBB can fallthrough itself, that can be changed
- // into a jmp.
- MachineBasicBlock *TBB = 0, *FBB = 0;
- SmallVector<MachineOperand, 4> Cond;
- if (TII->AnalyzeBranch(*SuccMBB, TBB, FBB, Cond))
- continue;
- if (!TBB && Cond.empty())
- TBB = next(MachineFunction::iterator(SuccMBB));
- else if (!FBB && !Cond.empty())
- FBB = next(MachineFunction::iterator(SuccMBB));
-
- // This calculate the cost of the transformation. Also, it finds the *only*
- // intra-loop edge if there is one.
- int Cost = 0;
- bool HasOneIntraSucc = true;
- MachineBasicBlock *IntraSucc = 0;
- for (MachineBasicBlock::succ_iterator SI = SuccMBB->succ_begin(),
- SE = SuccMBB->succ_end(); SI != SE; ++SI) {
- MachineBasicBlock *SSMBB = *SI;
- if (MLI->getLoopFor(SSMBB) == L) {
- if (!IntraSucc)
- IntraSucc = SSMBB;
- else
- HasOneIntraSucc = false;
+ // If the loop previously didn't exit with a fall-through and it now does,
+ // we eliminated a branch.
+ if (Changed &&
+ !BotHasFallthrough &&
+ HasFallthrough(L->getBottomBlock())) {
+ ++NumIntraElim;
+ }
+
+ return Changed;
+}
+
+/// MoveDiscontiguousLoopBlocks - Move any loop blocks that are not in the
+/// portion of the loop contiguous with the header. This usually makes the loop
+/// contiguous, provided that AnalyzeBranch can handle all the relevant
+/// branching. See the @cfg_islands case in test/CodeGen/X86/loop_blocks.ll
+/// for an example of this.
+bool CodePlacementOpt::MoveDiscontiguousLoopBlocks(MachineFunction &MF,
+ MachineLoop *L) {
+ bool Changed = false;
+ MachineBasicBlock *TopMBB = L->getTopBlock();
+ MachineBasicBlock *BotMBB = L->getBottomBlock();
+
+ // Determine a position to move orphaned loop blocks to. If TopMBB is not
+ // entered via fallthrough and BotMBB is exited via fallthrough, prepend them
+ // to the top of the loop to avoid loosing that fallthrough. Otherwise append
+ // them to the bottom, even if it previously had a fallthrough, on the theory
+ // that it's worth an extra branch to keep the loop contiguous.
+ MachineFunction::iterator InsertPt =
+ llvm::next(MachineFunction::iterator(BotMBB));
+ bool InsertAtTop = false;
+ if (TopMBB != MF.begin() &&
+ !HasFallthrough(prior(MachineFunction::iterator(TopMBB))) &&
+ HasFallthrough(BotMBB)) {
+ InsertPt = TopMBB;
+ InsertAtTop = true;
+ }
+
+ // Keep a record of which blocks are in the portion of the loop contiguous
+ // with the loop header.
+ SmallPtrSet<MachineBasicBlock *, 8> ContiguousBlocks;
+ for (MachineFunction::iterator I = TopMBB,
+ E = llvm::next(MachineFunction::iterator(BotMBB)); I != E; ++I)
+ ContiguousBlocks.insert(I);
+
+ // Find non-contigous blocks and fix them.
+ if (InsertPt != MF.begin() && HasAnalyzableTerminator(prior(InsertPt)))
+ for (MachineLoop::block_iterator BI = L->block_begin(), BE = L->block_end();
+ BI != BE; ++BI) {
+ MachineBasicBlock *BB = *BI;
+
+ // Verify that we can analyze all the loop entry edges before beginning
+ // any changes which will require us to be able to analyze them.
+ if (!HasAnalyzableTerminator(BB))
+ continue;
+ if (!HasAnalyzableTerminator(prior(MachineFunction::iterator(BB))))
+ continue;
+
+ // If the layout predecessor is part of the loop, this block will be
+ // processed along with it. This keeps them in their relative order.
+ if (BB != MF.begin() &&
+ L->contains(prior(MachineFunction::iterator(BB))))
+ continue;
+
+ // Check to see if this block is already contiguous with the main
+ // portion of the loop.
+ if (!ContiguousBlocks.insert(BB))
+ continue;
+
+ // Move the block.
+ DEBUG(dbgs() << "CGP: Moving blocks starting at BB#" << BB->getNumber()
+ << " to be contiguous with loop.\n");
+ Changed = true;
+
+ // Process this block and all loop blocks contiguous with it, to keep
+ // them in their relative order.
+ MachineFunction::iterator Begin = BB;
+ MachineFunction::iterator End = llvm::next(MachineFunction::iterator(BB));
+ for (; End != MF.end(); ++End) {
+ if (!L->contains(End)) break;
+ if (!HasAnalyzableTerminator(End)) break;
+ ContiguousBlocks.insert(End);
+ ++NumIntraMoved;
}
- if (SuccMBB->isLayoutSuccessor(SSMBB))
- // This will become a jmp.
- ++Cost;
- else if (MBB->isLayoutSuccessor(SSMBB))
- // One of the successor will become the new fallthrough.
- if (SSMBB == FBB) {
- FBB = 0;
- --Cost;
- } else if (!FBB && SSMBB == TBB && Cond.empty()) {
- TBB = 0;
- --Cost;
- } else if (!TII->ReverseBranchCondition(Cond)) {
- TBB = FBB;
- FBB = 0;
- --Cost;
+ // If we're inserting at the bottom of the loop, and the code we're
+ // moving originally had fall-through successors, bring the sucessors
+ // up with the loop blocks to preserve the fall-through edges.
+ if (!InsertAtTop)
+ for (; End != MF.end(); ++End) {
+ if (L->contains(End)) break;
+ if (!HasAnalyzableTerminator(End)) break;
+ if (!HasFallthrough(prior(End))) break;
}
- }
- if (Cost)
- continue;
-
- // Now, let's move the successor to below the BB to eliminate the jmp.
- SuccMBB->moveAfter(MBB);
- TII->RemoveBranch(*MBB);
- TII->RemoveBranch(*SuccMBB);
- if (TBB)
- TII->InsertBranch(*SuccMBB, TBB, FBB, Cond);
- ChangedMBBs.insert(MBB);
- ChangedMBBs.insert(SuccMBB);
- if (FtMBB) {
- TII->RemoveBranch(*FtMBB);
- TII->InsertBranch(*FtMBB, FtTBB, FtFBB, FtCond);
- ChangedMBBs.insert(FtMBB);
- }
- // If BB is the loop latch, we may have a new loop headr.
- if (MBB == L->getLoopLatch()) {
- assert(MLI->isLoopHeader(SuccMBB) &&
- "Only succ of loop latch is not the header?");
- if (HasOneIntraSucc && IntraSucc)
- std::replace(LoopHeaders.begin(),LoopHeaders.end(), SuccMBB, IntraSucc);
+ // Move the blocks. This may invalidate TopMBB and/or BotMBB, but
+ // we don't need them anymore at this point.
+ Splice(MF, InsertPt, Begin, End);
}
- }
- ++NumIntraMoved;
+ return Changed;
+}
+
+/// OptimizeIntraLoopEdgesInLoopNest - Reposition loop blocks to minimize
+/// intra-loop branching and to form contiguous loops.
+///
+/// This code takes the approach of making minor changes to the existing
+/// layout to fix specific loop-oriented problems. Also, it depends on
+/// AnalyzeBranch, which can't understand complex control instructions.
+///
+bool CodePlacementOpt::OptimizeIntraLoopEdgesInLoopNest(MachineFunction &MF,
+ MachineLoop *L) {
+ bool Changed = false;
+
+ // Do optimization for nested loops.
+ for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I)
+ Changed |= OptimizeIntraLoopEdgesInLoopNest(MF, *I);
+
+ // Do optimization for this loop.
+ Changed |= EliminateUnconditionalJumpsToTop(MF, L);
+ Changed |= MoveDiscontiguousLoopBlocks(MF, L);
+
+ return Changed;
+}
+
+/// OptimizeIntraLoopEdges - Reposition loop blocks to minimize
+/// intra-loop branching and to form contiguous loops.
+///
+bool CodePlacementOpt::OptimizeIntraLoopEdges(MachineFunction &MF) {
+ bool Changed = false;
+
+ if (!TLI->shouldOptimizeCodePlacement())
+ return Changed;
+
+ // Do optimization for each loop in the function.
+ for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end();
+ I != E; ++I)
+ if (!(*I)->getParentLoop())
+ Changed |= OptimizeIntraLoopEdgesInLoopNest(MF, *I);
+
return Changed;
}
if (!Align)
return false; // Don't care about loop alignment.
- // Make sure blocks are numbered in order
- MF.RenumberBlocks();
+ bool Changed = false;
+
+ for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end();
+ I != E; ++I)
+ Changed |= AlignLoop(MF, *I, Align);
+
+ return Changed;
+}
+/// AlignLoop - Align loop headers to target preferred alignments.
+///
+bool CodePlacementOpt::AlignLoop(MachineFunction &MF, MachineLoop *L,
+ unsigned Align) {
bool Changed = false;
- for (unsigned i = 0, e = LoopHeaders.size(); i != e; ++i) {
- MachineBasicBlock *HeaderMBB = LoopHeaders[i];
- MachineBasicBlock *PredMBB = prior(MachineFunction::iterator(HeaderMBB));
- if (MLI->getLoopFor(HeaderMBB) != MLI->getLoopFor(PredMBB)) {
- // If previously BB is in the same loop, don't align this BB. We want
- // to prevent adding noop's inside a loop.
- HeaderMBB->setAlignment(Align);
- Changed = true;
- ++NumHeaderAligned;
- }
- }
+
+ // Do alignment for nested loops.
+ for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I)
+ Changed |= AlignLoop(MF, *I, Align);
+
+ L->getTopBlock()->setAlignment(Align);
+ Changed = true;
+ ++NumLoopsAligned;
return Changed;
}
TLI = MF.getTarget().getTargetLowering();
TII = MF.getTarget().getInstrInfo();
- // Analyze the BBs first and keep track of loop headers and BBs that
- // end with an unconditional jmp to another block in the same loop.
- for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
- MachineBasicBlock *MBB = I;
- if (MBB->isLandingPad())
- continue;
- MachineLoop *L = MLI->getLoopFor(MBB);
- if (!L)
- continue;
- if (MLI->isLoopHeader(MBB))
- LoopHeaders.push_back(MBB);
-
- MachineBasicBlock *TBB = 0, *FBB = 0;
- SmallVector<MachineOperand, 4> Cond;
- if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond) || !Cond.empty())
- continue;
- if (MLI->getLoopFor(TBB) == L && !TBB->isLandingPad())
- UncondJmpMBBs.push_back(std::make_pair(MBB, TBB));
- }
-
- bool Changed = OptimizeIntraLoopEdges();
+ bool Changed = OptimizeIntraLoopEdges(MF);
Changed |= AlignLoops(MF);
- ChangedMBBs.clear();
- UncondJmpMBBs.clear();
- LoopHeaders.clear();
-
return Changed;
}