+ 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) {
+ unsigned Time = 0;
+ for (; I != E; ++I) {
+ const TargetInstrDesc &TID = I->getDesc();
+ if (TID.isCall())
+ Time += 10;
+ else if (TID.isSimpleLoad() || TID.mayStore())
+ 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,
+ MachineBasicBlock *PredBB) {
+ // If one block is the entry block, split the other one; we can't generate
+ // a branch to the entry block, as its label is not emitted.
+ MachineBasicBlock *Entry = MBB1->getParent()->begin();
+ if (MBB1 == Entry)
+ return false;
+ if (MBB2 == Entry)
+ return true;
+
+ // If one block falls through into the common successor, choose that
+ // one to split; it is one instruction less to do that.
+ if (PredBB) {
+ if (MBB1 == PredBB)
+ return true;
+ else if (MBB2 == PredBB)
+ return false;
+ }
+ // 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);
+ unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I);
+
+ // 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;
+}
+
+// CurMBB needs to add an unconditional branch to SuccMBB (we removed these
+// branches temporarily for tail merging). In the case where CurMBB ends
+// with a conditional branch to the next block, optimize by reversing the
+// test and conditionally branching to SuccMBB instead.
+
+static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB,
+ const TargetInstrInfo *TII) {
+ MachineFunction *MF = CurMBB->getParent();
+ MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB));
+ MachineBasicBlock *TBB = 0, *FBB = 0;
+ std::vector<MachineOperand> Cond;
+ if (I != MF->end() &&
+ !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond)) {
+ MachineBasicBlock *NextBB = I;
+ if (TBB == NextBB && Cond.size() && !FBB) {
+ if (!TII->ReverseBranchCondition(Cond)) {
+ TII->RemoveBranch(*CurMBB);
+ TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
+ return;
+ }
+ }
+ }
+ TII->InsertBranch(*CurMBB, SuccBB, NULL, std::vector<MachineOperand>());
+}
+
+static bool MergeCompare(const std::pair<unsigned,MachineBasicBlock*> &p,
+ const std::pair<unsigned,MachineBasicBlock*> &q) {
+ if (p.first < q.first)
+ return true;
+ else if (p.first > q.first)
+ return false;
+ else if (p.second->getNumber() < q.second->getNumber())
+ return true;
+ else if (p.second->getNumber() > q.second->getNumber())
+ return false;
+ else {
+ // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
+ // an object with itself.
+#ifndef _GLIBCXX_DEBUG
+ assert(0 && "Predecessor appears twice");
+#endif
+ return(false);
+ }
+}
+
+// See if any of the blocks in MergePotentials (which all have a common single
+// successor, or all have no successor) can be tail-merged. If there is a
+// successor, any blocks in MergePotentials that are not tail-merged and
+// are not immediately before Succ must have an unconditional branch to
+// Succ added (but the predecessor/successor lists need no adjustment).
+// The lone predecessor of Succ that falls through into Succ,
+// if any, is given in PredBB.
+
+bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB,
+ MachineBasicBlock* PredBB) {
+ unsigned minCommonTailLength = (SuccBB ? 1 : 2);
+ MadeChange = false;
+
+ // Sort by hash value so that blocks with identical end sequences sort
+ // together.
+ std::stable_sort(MergePotentials.begin(), MergePotentials.end(), MergeCompare);
+
+ // Walk through equivalence sets looking for actual exact matches.
+ while (MergePotentials.size() > 1) {
+ unsigned CurHash = (MergePotentials.end()-1)->first;
+ unsigned PrevHash = (MergePotentials.end()-2)->first;
+ MachineBasicBlock *CurMBB = (MergePotentials.end()-1)->second;
+
+ // If there is nothing that matches the hash of the current basic block,
+ // give up.
+ if (CurHash != PrevHash) {
+ if (SuccBB && CurMBB != PredBB)
+ FixTail(CurMBB, SuccBB, TII);
+ MergePotentials.pop_back();
+ continue;
+ }
+
+ // Look through all the pairs of blocks that have the same hash as this
+ // one, and find the pair that has the largest number of instructions in
+ // common.
+ // Since instructions may get combined later (e.g. single stores into
+ // store multiple) this measure is not particularly accurate.
+ MachineBasicBlock::iterator BBI1, BBI2;
+
+ unsigned FoundI = ~0U, FoundJ = ~0U;
+ unsigned maxCommonTailLength = 0U;
+ for (int i = MergePotentials.size()-1;
+ i != -1 && MergePotentials[i].first == CurHash; --i) {
+ for (int j = i-1;
+ j != -1 && MergePotentials[j].first == CurHash; --j) {
+ MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
+ unsigned CommonTailLen = ComputeCommonTailLength(
+ MergePotentials[i].second,
+ MergePotentials[j].second,
+ TrialBBI1, TrialBBI2);
+ if (CommonTailLen >= minCommonTailLength &&
+ CommonTailLen > maxCommonTailLength) {
+ FoundI = i;
+ FoundJ = j;
+ maxCommonTailLength = CommonTailLen;
+ BBI1 = TrialBBI1;
+ BBI2 = TrialBBI2;
+ }
+ }
+ }
+
+ // If we didn't find any pair that has at least minCommonTailLength
+ // instructions in common, bail out. All entries with this
+ // hash code can go away now.
+ if (FoundI == ~0U) {
+ for (int i = MergePotentials.size()-1;
+ i != -1 && MergePotentials[i].first == CurHash; --i) {
+ // Put the unconditional branch back, if we need one.
+ CurMBB = MergePotentials[i].second;
+ if (SuccBB && CurMBB != PredBB)
+ FixTail(CurMBB, SuccBB, TII);
+ MergePotentials.pop_back();
+ }
+ continue;
+ }
+
+ // Otherwise, move the block(s) to the right position(s). So that
+ // BBI1/2 will be valid, the last must be I and the next-to-last J.
+ if (FoundI != MergePotentials.size()-1)
+ std::swap(MergePotentials[FoundI], *(MergePotentials.end()-1));
+ if (FoundJ != MergePotentials.size()-2)
+ std::swap(MergePotentials[FoundJ], *(MergePotentials.end()-2));
+
+ CurMBB = (MergePotentials.end()-1)->second;
+ MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second;
+
+ // If neither block is the entire common tail, split the tail of one block
+ // to make it redundant with the other tail. Also, we cannot jump to the
+ // entry block, so if one block is the entry block, split the other one.
+ MachineBasicBlock *Entry = CurMBB->getParent()->begin();
+ if (CurMBB->begin() == BBI1 && CurMBB != Entry)
+ ; // CurMBB is common tail
+ else if (MBB2->begin() == BBI2 && MBB2 != Entry)
+ ; // MBB2 is common tail
+ else {
+ 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, PredBB)) {
+ 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 && MBB2 != Entry) {
+ // 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();
+ } else {
+ assert(CurMBB->begin() == BBI1 && CurMBB != Entry &&
+ "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);
+ }
+ MadeChange = true;
+ }
+ return MadeChange;