Home | History | Annotate | Download | only in CodeGen
      1 //===-- CodePlacementOpt.cpp - Code Placement pass. -----------------------===//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 //
     10 // This file implements the pass that optimizes code placement and aligns loop
     11 // headers to target-specific alignment boundaries.
     12 //
     13 //===----------------------------------------------------------------------===//
     14 
     15 #define DEBUG_TYPE "code-placement"
     16 #include "llvm/CodeGen/Passes.h"
     17 #include "llvm/ADT/Statistic.h"
     18 #include "llvm/CodeGen/MachineFunctionPass.h"
     19 #include "llvm/CodeGen/MachineLoopInfo.h"
     20 #include "llvm/Support/Compiler.h"
     21 #include "llvm/Support/Debug.h"
     22 #include "llvm/Target/TargetInstrInfo.h"
     23 #include "llvm/Target/TargetLowering.h"
     24 #include "llvm/Target/TargetMachine.h"
     25 using namespace llvm;
     26 
     27 STATISTIC(NumLoopsAligned,  "Number of loops aligned");
     28 STATISTIC(NumIntraElim,     "Number of intra loop branches eliminated");
     29 STATISTIC(NumIntraMoved,    "Number of intra loop branches moved");
     30 
     31 namespace {
     32   class CodePlacementOpt : public MachineFunctionPass {
     33     const MachineLoopInfo *MLI;
     34     const TargetInstrInfo *TII;
     35     const TargetLowering  *TLI;
     36 
     37   public:
     38     static char ID;
     39     CodePlacementOpt() : MachineFunctionPass(ID) {}
     40 
     41     virtual bool runOnMachineFunction(MachineFunction &MF);
     42 
     43     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
     44       AU.addRequired<MachineLoopInfo>();
     45       AU.addPreservedID(MachineDominatorsID);
     46       MachineFunctionPass::getAnalysisUsage(AU);
     47     }
     48 
     49   private:
     50     bool HasFallthrough(MachineBasicBlock *MBB);
     51     bool HasAnalyzableTerminator(MachineBasicBlock *MBB);
     52     void Splice(MachineFunction &MF,
     53                 MachineFunction::iterator InsertPt,
     54                 MachineFunction::iterator Begin,
     55                 MachineFunction::iterator End);
     56     bool EliminateUnconditionalJumpsToTop(MachineFunction &MF,
     57                                           MachineLoop *L);
     58     bool MoveDiscontiguousLoopBlocks(MachineFunction &MF,
     59                                      MachineLoop *L);
     60     bool OptimizeIntraLoopEdgesInLoopNest(MachineFunction &MF, MachineLoop *L);
     61     bool OptimizeIntraLoopEdges(MachineFunction &MF);
     62     bool AlignLoops(MachineFunction &MF);
     63     bool AlignLoop(MachineFunction &MF, MachineLoop *L, unsigned Align);
     64   };
     65 
     66   char CodePlacementOpt::ID = 0;
     67 } // end anonymous namespace
     68 
     69 char &llvm::CodePlacementOptID = CodePlacementOpt::ID;
     70 INITIALIZE_PASS(CodePlacementOpt, "code-placement",
     71                 "Code Placement Optimizer", false, false)
     72 
     73 /// HasFallthrough - Test whether the given branch has a fallthrough, either as
     74 /// a plain fallthrough or as a fallthrough case of a conditional branch.
     75 ///
     76 bool CodePlacementOpt::HasFallthrough(MachineBasicBlock *MBB) {
     77   MachineBasicBlock *TBB = 0, *FBB = 0;
     78   SmallVector<MachineOperand, 4> Cond;
     79   if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond))
     80     return false;
     81   // This conditional branch has no fallthrough.
     82   if (FBB)
     83     return false;
     84   // An unconditional branch has no fallthrough.
     85   if (Cond.empty() && TBB)
     86     return false;
     87   // It has a fallthrough.
     88   return true;
     89 }
     90 
     91 /// HasAnalyzableTerminator - Test whether AnalyzeBranch will succeed on MBB.
     92 /// This is called before major changes are begun to test whether it will be
     93 /// possible to complete the changes.
     94 ///
     95 /// Target-specific code is hereby encouraged to make AnalyzeBranch succeed
     96 /// whenever possible.
     97 ///
     98 bool CodePlacementOpt::HasAnalyzableTerminator(MachineBasicBlock *MBB) {
     99   // Conservatively ignore EH landing pads.
    100   if (MBB->isLandingPad()) return false;
    101 
    102   // Aggressively handle return blocks and similar constructs.
    103   if (MBB->succ_empty()) return true;
    104 
    105   // Ask the target's AnalyzeBranch if it can handle this block.
    106   MachineBasicBlock *TBB = 0, *FBB = 0;
    107   SmallVector<MachineOperand, 4> Cond;
    108   // Make sure the terminator is understood.
    109   if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond))
    110     return false;
    111    // Ignore blocks which look like they might have EH-related control flow.
    112    // AnalyzeBranch thinks it knows how to analyze such things, but it doesn't
    113    // recognize the possibility of a control transfer through an unwind.
    114    // Such blocks contain EH_LABEL instructions, however they may be in the
    115    // middle of the block. Instead of searching for them, just check to see
    116    // if the CFG disagrees with AnalyzeBranch.
    117   if (1u + !Cond.empty() != MBB->succ_size())
    118     return false;
    119   // Make sure we have the option of reversing the condition.
    120   if (!Cond.empty() && TII->ReverseBranchCondition(Cond))
    121     return false;
    122   return true;
    123 }
    124 
    125 /// Splice - Move the sequence of instructions [Begin,End) to just before
    126 /// InsertPt. Update branch instructions as needed to account for broken
    127 /// fallthrough edges and to take advantage of newly exposed fallthrough
    128 /// opportunities.
    129 ///
    130 void CodePlacementOpt::Splice(MachineFunction &MF,
    131                               MachineFunction::iterator InsertPt,
    132                               MachineFunction::iterator Begin,
    133                               MachineFunction::iterator End) {
    134   assert(Begin != MF.begin() && End != MF.begin() && InsertPt != MF.begin() &&
    135          "Splice can't change the entry block!");
    136   MachineFunction::iterator OldBeginPrior = prior(Begin);
    137   MachineFunction::iterator OldEndPrior = prior(End);
    138 
    139   MF.splice(InsertPt, Begin, End);
    140 
    141   prior(Begin)->updateTerminator();
    142   OldBeginPrior->updateTerminator();
    143   OldEndPrior->updateTerminator();
    144 }
    145 
    146 /// EliminateUnconditionalJumpsToTop - Move blocks which unconditionally jump
    147 /// to the loop top to the top of the loop so that they have a fall through.
    148 /// This can introduce a branch on entry to the loop, but it can eliminate a
    149 /// branch within the loop. See the @simple case in
    150 /// test/CodeGen/X86/loop_blocks.ll for an example of this.
    151 bool CodePlacementOpt::EliminateUnconditionalJumpsToTop(MachineFunction &MF,
    152                                                         MachineLoop *L) {
    153   bool Changed = false;
    154   MachineBasicBlock *TopMBB = L->getTopBlock();
    155 
    156   bool BotHasFallthrough = HasFallthrough(L->getBottomBlock());
    157 
    158   if (TopMBB == MF.begin() ||
    159       HasAnalyzableTerminator(prior(MachineFunction::iterator(TopMBB)))) {
    160   new_top:
    161     for (MachineBasicBlock::pred_iterator PI = TopMBB->pred_begin(),
    162          PE = TopMBB->pred_end(); PI != PE; ++PI) {
    163       MachineBasicBlock *Pred = *PI;
    164       if (Pred == TopMBB) continue;
    165       if (HasFallthrough(Pred)) continue;
    166       if (!L->contains(Pred)) continue;
    167 
    168       // Verify that we can analyze all the loop entry edges before beginning
    169       // any changes which will require us to be able to analyze them.
    170       if (Pred == MF.begin())
    171         continue;
    172       if (!HasAnalyzableTerminator(Pred))
    173         continue;
    174       if (!HasAnalyzableTerminator(prior(MachineFunction::iterator(Pred))))
    175         continue;
    176 
    177       // Move the block.
    178       DEBUG(dbgs() << "CGP: Moving blocks starting at BB#" << Pred->getNumber()
    179                    << " to top of loop.\n");
    180       Changed = true;
    181 
    182       // Move it and all the blocks that can reach it via fallthrough edges
    183       // exclusively, to keep existing fallthrough edges intact.
    184       MachineFunction::iterator Begin = Pred;
    185       MachineFunction::iterator End = llvm::next(Begin);
    186       while (Begin != MF.begin()) {
    187         MachineFunction::iterator Prior = prior(Begin);
    188         if (Prior == MF.begin())
    189           break;
    190         // Stop when a non-fallthrough edge is found.
    191         if (!HasFallthrough(Prior))
    192           break;
    193         // Stop if a block which could fall-through out of the loop is found.
    194         if (Prior->isSuccessor(End))
    195           break;
    196         // If we've reached the top, stop scanning.
    197         if (Prior == MachineFunction::iterator(TopMBB)) {
    198           // We know top currently has a fall through (because we just checked
    199           // it) which would be lost if we do the transformation, so it isn't
    200           // worthwhile to do the transformation unless it would expose a new
    201           // fallthrough edge.
    202           if (!Prior->isSuccessor(End))
    203             goto next_pred;
    204           // Otherwise we can stop scanning and proceed to move the blocks.
    205           break;
    206         }
    207         // If we hit a switch or something complicated, don't move anything
    208         // for this predecessor.
    209         if (!HasAnalyzableTerminator(prior(MachineFunction::iterator(Prior))))
    210           break;
    211         // Ok, the block prior to Begin will be moved along with the rest.
    212         // Extend the range to include it.
    213         Begin = Prior;
    214         ++NumIntraMoved;
    215       }
    216 
    217       // Move the blocks.
    218       Splice(MF, TopMBB, Begin, End);
    219 
    220       // Update TopMBB.
    221       TopMBB = L->getTopBlock();
    222 
    223       // We have a new loop top. Iterate on it. We shouldn't have to do this
    224       // too many times if BranchFolding has done a reasonable job.
    225       goto new_top;
    226     next_pred:;
    227     }
    228   }
    229 
    230   // If the loop previously didn't exit with a fall-through and it now does,
    231   // we eliminated a branch.
    232   if (Changed &&
    233       !BotHasFallthrough &&
    234       HasFallthrough(L->getBottomBlock())) {
    235     ++NumIntraElim;
    236   }
    237 
    238   return Changed;
    239 }
    240 
    241 /// MoveDiscontiguousLoopBlocks - Move any loop blocks that are not in the
    242 /// portion of the loop contiguous with the header. This usually makes the loop
    243 /// contiguous, provided that AnalyzeBranch can handle all the relevant
    244 /// branching. See the @cfg_islands case in test/CodeGen/X86/loop_blocks.ll
    245 /// for an example of this.
    246 bool CodePlacementOpt::MoveDiscontiguousLoopBlocks(MachineFunction &MF,
    247                                                    MachineLoop *L) {
    248   bool Changed = false;
    249   MachineBasicBlock *TopMBB = L->getTopBlock();
    250   MachineBasicBlock *BotMBB = L->getBottomBlock();
    251 
    252   // Determine a position to move orphaned loop blocks to. If TopMBB is not
    253   // entered via fallthrough and BotMBB is exited via fallthrough, prepend them
    254   // to the top of the loop to avoid losing that fallthrough. Otherwise append
    255   // them to the bottom, even if it previously had a fallthrough, on the theory
    256   // that it's worth an extra branch to keep the loop contiguous.
    257   MachineFunction::iterator InsertPt =
    258     llvm::next(MachineFunction::iterator(BotMBB));
    259   bool InsertAtTop = false;
    260   if (TopMBB != MF.begin() &&
    261       !HasFallthrough(prior(MachineFunction::iterator(TopMBB))) &&
    262       HasFallthrough(BotMBB)) {
    263     InsertPt = TopMBB;
    264     InsertAtTop = true;
    265   }
    266 
    267   // Keep a record of which blocks are in the portion of the loop contiguous
    268   // with the loop header.
    269   SmallPtrSet<MachineBasicBlock *, 8> ContiguousBlocks;
    270   for (MachineFunction::iterator I = TopMBB,
    271        E = llvm::next(MachineFunction::iterator(BotMBB)); I != E; ++I)
    272     ContiguousBlocks.insert(I);
    273 
    274   // Find non-contigous blocks and fix them.
    275   if (InsertPt != MF.begin() && HasAnalyzableTerminator(prior(InsertPt)))
    276     for (MachineLoop::block_iterator BI = L->block_begin(), BE = L->block_end();
    277          BI != BE; ++BI) {
    278       MachineBasicBlock *BB = *BI;
    279 
    280       // Verify that we can analyze all the loop entry edges before beginning
    281       // any changes which will require us to be able to analyze them.
    282       if (!HasAnalyzableTerminator(BB))
    283         continue;
    284       if (!HasAnalyzableTerminator(prior(MachineFunction::iterator(BB))))
    285         continue;
    286 
    287       // If the layout predecessor is part of the loop, this block will be
    288       // processed along with it. This keeps them in their relative order.
    289       if (BB != MF.begin() &&
    290           L->contains(prior(MachineFunction::iterator(BB))))
    291         continue;
    292 
    293       // Check to see if this block is already contiguous with the main
    294       // portion of the loop.
    295       if (!ContiguousBlocks.insert(BB))
    296         continue;
    297 
    298       // Move the block.
    299       DEBUG(dbgs() << "CGP: Moving blocks starting at BB#" << BB->getNumber()
    300                    << " to be contiguous with loop.\n");
    301       Changed = true;
    302 
    303       // Process this block and all loop blocks contiguous with it, to keep
    304       // them in their relative order.
    305       MachineFunction::iterator Begin = BB;
    306       MachineFunction::iterator End = llvm::next(MachineFunction::iterator(BB));
    307       for (; End != MF.end(); ++End) {
    308         if (!L->contains(End)) break;
    309         if (!HasAnalyzableTerminator(End)) break;
    310         ContiguousBlocks.insert(End);
    311         ++NumIntraMoved;
    312       }
    313 
    314       // If we're inserting at the bottom of the loop, and the code we're
    315       // moving originally had fall-through successors, bring the sucessors
    316       // up with the loop blocks to preserve the fall-through edges.
    317       if (!InsertAtTop)
    318         for (; End != MF.end(); ++End) {
    319           if (L->contains(End)) break;
    320           if (!HasAnalyzableTerminator(End)) break;
    321           if (!HasFallthrough(prior(End))) break;
    322         }
    323 
    324       // Move the blocks. This may invalidate TopMBB and/or BotMBB, but
    325       // we don't need them anymore at this point.
    326       Splice(MF, InsertPt, Begin, End);
    327     }
    328 
    329   return Changed;
    330 }
    331 
    332 /// OptimizeIntraLoopEdgesInLoopNest - Reposition loop blocks to minimize
    333 /// intra-loop branching and to form contiguous loops.
    334 ///
    335 /// This code takes the approach of making minor changes to the existing
    336 /// layout to fix specific loop-oriented problems. Also, it depends on
    337 /// AnalyzeBranch, which can't understand complex control instructions.
    338 ///
    339 bool CodePlacementOpt::OptimizeIntraLoopEdgesInLoopNest(MachineFunction &MF,
    340                                                         MachineLoop *L) {
    341   bool Changed = false;
    342 
    343   // Do optimization for nested loops.
    344   for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I)
    345     Changed |= OptimizeIntraLoopEdgesInLoopNest(MF, *I);
    346 
    347   // Do optimization for this loop.
    348   Changed |= EliminateUnconditionalJumpsToTop(MF, L);
    349   Changed |= MoveDiscontiguousLoopBlocks(MF, L);
    350 
    351   return Changed;
    352 }
    353 
    354 /// OptimizeIntraLoopEdges - Reposition loop blocks to minimize
    355 /// intra-loop branching and to form contiguous loops.
    356 ///
    357 bool CodePlacementOpt::OptimizeIntraLoopEdges(MachineFunction &MF) {
    358   bool Changed = false;
    359 
    360   if (!TLI->shouldOptimizeCodePlacement())
    361     return Changed;
    362 
    363   // Do optimization for each loop in the function.
    364   for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end();
    365        I != E; ++I)
    366     if (!(*I)->getParentLoop())
    367       Changed |= OptimizeIntraLoopEdgesInLoopNest(MF, *I);
    368 
    369   return Changed;
    370 }
    371 
    372 /// AlignLoops - Align loop headers to target preferred alignments.
    373 ///
    374 bool CodePlacementOpt::AlignLoops(MachineFunction &MF) {
    375   const Function *F = MF.getFunction();
    376   if (F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
    377                                       Attribute::OptimizeForSize))
    378     return false;
    379 
    380   unsigned Align = TLI->getPrefLoopAlignment();
    381   if (!Align)
    382     return false;  // Don't care about loop alignment.
    383 
    384   bool Changed = false;
    385 
    386   for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end();
    387        I != E; ++I)
    388     Changed |= AlignLoop(MF, *I, Align);
    389 
    390   return Changed;
    391 }
    392 
    393 /// AlignLoop - Align loop headers to target preferred alignments.
    394 ///
    395 bool CodePlacementOpt::AlignLoop(MachineFunction &MF, MachineLoop *L,
    396                                  unsigned Align) {
    397   bool Changed = false;
    398 
    399   // Do alignment for nested loops.
    400   for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I)
    401     Changed |= AlignLoop(MF, *I, Align);
    402 
    403   L->getTopBlock()->setAlignment(Align);
    404   Changed = true;
    405   ++NumLoopsAligned;
    406 
    407   return Changed;
    408 }
    409 
    410 bool CodePlacementOpt::runOnMachineFunction(MachineFunction &MF) {
    411   MLI = &getAnalysis<MachineLoopInfo>();
    412   if (MLI->empty())
    413     return false;  // No loops.
    414 
    415   TLI = MF.getTarget().getTargetLowering();
    416   TII = MF.getTarget().getInstrInfo();
    417 
    418   bool Changed = OptimizeIntraLoopEdges(MF);
    419 
    420   Changed |= AlignLoops(MF);
    421 
    422   return Changed;
    423 }
    424