MachineSink.cpp revision a858f3e98f4ebe01e1e9fb3f1733dbcd4799d452
1//===-- MachineSink.cpp - Sinking for machine instructions ----------------===//
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 pass moves instructions into successor blocks, when possible, so that
11// they aren't executed on paths where their results aren't needed.
12//
13// This pass is not intended to be a replacement or a complete alternative
14// for an LLVM-IR-level sinking pass. It is only designed to sink simple
15// constructs that are not exposed before lowering and instruction selection.
16//
17//===----------------------------------------------------------------------===//
18
19#define DEBUG_TYPE "machine-sink"
20#include "llvm/CodeGen/Passes.h"
21#include "llvm/CodeGen/MachineRegisterInfo.h"
22#include "llvm/CodeGen/MachineDominators.h"
23#include "llvm/CodeGen/MachineLoopInfo.h"
24#include "llvm/Analysis/AliasAnalysis.h"
25#include "llvm/Target/TargetRegisterInfo.h"
26#include "llvm/Target/TargetInstrInfo.h"
27#include "llvm/Target/TargetMachine.h"
28#include "llvm/ADT/Statistic.h"
29#include "llvm/Support/Debug.h"
30#include "llvm/Support/raw_ostream.h"
31using namespace llvm;
32
33STATISTIC(NumSunk, "Number of machine instructions sunk");
34
35namespace {
36  class MachineSinking : public MachineFunctionPass {
37    const TargetInstrInfo *TII;
38    const TargetRegisterInfo *TRI;
39    MachineRegisterInfo  *RegInfo; // Machine register information
40    MachineDominatorTree *DT;   // Machine dominator tree
41    MachineLoopInfo *LI;
42    AliasAnalysis *AA;
43    BitVector AllocatableSet;   // Which physregs are allocatable?
44
45  public:
46    static char ID; // Pass identification
47    MachineSinking() : MachineFunctionPass(&ID) {}
48
49    virtual bool runOnMachineFunction(MachineFunction &MF);
50
51    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
52      AU.setPreservesCFG();
53      MachineFunctionPass::getAnalysisUsage(AU);
54      AU.addRequired<AliasAnalysis>();
55      AU.addRequired<MachineDominatorTree>();
56      AU.addRequired<MachineLoopInfo>();
57      AU.addPreserved<MachineDominatorTree>();
58      AU.addPreserved<MachineLoopInfo>();
59    }
60  private:
61    bool ProcessBlock(MachineBasicBlock &MBB);
62    bool SinkInstruction(MachineInstr *MI, bool &SawStore);
63    bool AllUsesDominatedByBlock(unsigned Reg, MachineBasicBlock *MBB) const;
64  };
65} // end anonymous namespace
66
67char MachineSinking::ID = 0;
68static RegisterPass<MachineSinking>
69X("machine-sink", "Machine code sinking");
70
71FunctionPass *llvm::createMachineSinkingPass() { return new MachineSinking(); }
72
73/// AllUsesDominatedByBlock - Return true if all uses of the specified register
74/// occur in blocks dominated by the specified block.
75bool MachineSinking::AllUsesDominatedByBlock(unsigned Reg,
76                                             MachineBasicBlock *MBB) const {
77  assert(TargetRegisterInfo::isVirtualRegister(Reg) &&
78         "Only makes sense for vregs");
79  // Ignoring debug uses is necessary so debug info doesn't affect the code.
80  // This may leave a referencing dbg_value in the original block, before
81  // the definition of the vreg.  Dwarf generator handles this although the
82  // user might not get the right info at runtime.
83  for (MachineRegisterInfo::use_nodbg_iterator I =
84       RegInfo->use_nodbg_begin(Reg),
85       E = RegInfo->use_nodbg_end(); I != E; ++I) {
86    // Determine the block of the use.
87    MachineInstr *UseInst = &*I;
88    MachineBasicBlock *UseBlock = UseInst->getParent();
89    if (UseInst->isPHI()) {
90      // PHI nodes use the operand in the predecessor block, not the block with
91      // the PHI.
92      UseBlock = UseInst->getOperand(I.getOperandNo()+1).getMBB();
93    }
94    // Check that it dominates.
95    if (!DT->dominates(MBB, UseBlock))
96      return false;
97  }
98  return true;
99}
100
101bool MachineSinking::runOnMachineFunction(MachineFunction &MF) {
102  DEBUG(dbgs() << "******** Machine Sinking ********\n");
103
104  const TargetMachine &TM = MF.getTarget();
105  TII = TM.getInstrInfo();
106  TRI = TM.getRegisterInfo();
107  RegInfo = &MF.getRegInfo();
108  DT = &getAnalysis<MachineDominatorTree>();
109  LI = &getAnalysis<MachineLoopInfo>();
110  AA = &getAnalysis<AliasAnalysis>();
111  AllocatableSet = TRI->getAllocatableSet(MF);
112
113  bool EverMadeChange = false;
114
115  while (1) {
116    bool MadeChange = false;
117
118    // Process all basic blocks.
119    for (MachineFunction::iterator I = MF.begin(), E = MF.end();
120         I != E; ++I)
121      MadeChange |= ProcessBlock(*I);
122
123    // If this iteration over the code changed anything, keep iterating.
124    if (!MadeChange) break;
125    EverMadeChange = true;
126  }
127  return EverMadeChange;
128}
129
130bool MachineSinking::ProcessBlock(MachineBasicBlock &MBB) {
131  // Can't sink anything out of a block that has less than two successors.
132  if (MBB.succ_size() <= 1 || MBB.empty()) return false;
133
134  // Don't bother sinking code out of unreachable blocks. In addition to being
135  // unprofitable, it can also lead to infinite looping, because in an unreachable
136  // loop there may be nowhere to stop.
137  if (!DT->isReachableFromEntry(&MBB)) return false;
138
139  bool MadeChange = false;
140
141  // Walk the basic block bottom-up.  Remember if we saw a store.
142  MachineBasicBlock::iterator I = MBB.end();
143  --I;
144  bool ProcessedBegin, SawStore = false;
145  do {
146    MachineInstr *MI = I;  // The instruction to sink.
147
148    // Predecrement I (if it's not begin) so that it isn't invalidated by
149    // sinking.
150    ProcessedBegin = I == MBB.begin();
151    if (!ProcessedBegin)
152      --I;
153
154    if (MI->isDebugValue())
155      continue;
156
157    if (SinkInstruction(MI, SawStore))
158      ++NumSunk, MadeChange = true;
159
160    // If we just processed the first instruction in the block, we're done.
161  } while (!ProcessedBegin);
162
163  return MadeChange;
164}
165
166/// SinkInstruction - Determine whether it is safe to sink the specified machine
167/// instruction out of its current block into a successor.
168bool MachineSinking::SinkInstruction(MachineInstr *MI, bool &SawStore) {
169  // Check if it's safe to move the instruction.
170  if (!MI->isSafeToMove(TII, AA, SawStore))
171    return false;
172
173  // FIXME: This should include support for sinking instructions within the
174  // block they are currently in to shorten the live ranges.  We often get
175  // instructions sunk into the top of a large block, but it would be better to
176  // also sink them down before their first use in the block.  This xform has to
177  // be careful not to *increase* register pressure though, e.g. sinking
178  // "x = y + z" down if it kills y and z would increase the live ranges of y
179  // and z and only shrink the live range of x.
180
181  // Loop over all the operands of the specified instruction.  If there is
182  // anything we can't handle, bail out.
183  MachineBasicBlock *ParentBlock = MI->getParent();
184
185  // SuccToSinkTo - This is the successor to sink this instruction to, once we
186  // decide.
187  MachineBasicBlock *SuccToSinkTo = 0;
188
189  for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
190    const MachineOperand &MO = MI->getOperand(i);
191    if (!MO.isReg()) continue;  // Ignore non-register operands.
192
193    unsigned Reg = MO.getReg();
194    if (Reg == 0) continue;
195
196    if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
197      if (MO.isUse()) {
198        // If the physreg has no defs anywhere, it's just an ambient register
199        // and we can freely move its uses. Alternatively, if it's allocatable,
200        // it could get allocated to something with a def during allocation.
201        if (!RegInfo->def_empty(Reg))
202          return false;
203        if (AllocatableSet.test(Reg))
204          return false;
205        // Check for a def among the register's aliases too.
206        for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
207          unsigned AliasReg = *Alias;
208          if (!RegInfo->def_empty(AliasReg))
209            return false;
210          if (AllocatableSet.test(AliasReg))
211            return false;
212        }
213      } else if (!MO.isDead()) {
214        // A def that isn't dead. We can't move it.
215        return false;
216      }
217    } else {
218      // Virtual register uses are always safe to sink.
219      if (MO.isUse()) continue;
220
221      // If it's not safe to move defs of the register class, then abort.
222      if (!TII->isSafeToMoveRegClassDefs(RegInfo->getRegClass(Reg)))
223        return false;
224
225      // FIXME: This picks a successor to sink into based on having one
226      // successor that dominates all the uses.  However, there are cases where
227      // sinking can happen but where the sink point isn't a successor.  For
228      // example:
229      //   x = computation
230      //   if () {} else {}
231      //   use x
232      // the instruction could be sunk over the whole diamond for the
233      // if/then/else (or loop, etc), allowing it to be sunk into other blocks
234      // after that.
235
236      // Virtual register defs can only be sunk if all their uses are in blocks
237      // dominated by one of the successors.
238      if (SuccToSinkTo) {
239        // If a previous operand picked a block to sink to, then this operand
240        // must be sinkable to the same block.
241        if (!AllUsesDominatedByBlock(Reg, SuccToSinkTo))
242          return false;
243        continue;
244      }
245
246      // Otherwise, we should look at all the successors and decide which one
247      // we should sink to.
248      for (MachineBasicBlock::succ_iterator SI = ParentBlock->succ_begin(),
249           E = ParentBlock->succ_end(); SI != E; ++SI) {
250        if (AllUsesDominatedByBlock(Reg, *SI)) {
251          SuccToSinkTo = *SI;
252          break;
253        }
254      }
255
256      // If we couldn't find a block to sink to, ignore this instruction.
257      if (SuccToSinkTo == 0)
258        return false;
259    }
260  }
261
262  // If there are no outputs, it must have side-effects.
263  if (SuccToSinkTo == 0)
264    return false;
265
266  // It's not safe to sink instructions to EH landing pad. Control flow into
267  // landing pad is implicitly defined.
268  if (SuccToSinkTo->isLandingPad())
269    return false;
270
271  // It is not possible to sink an instruction into its own block.  This can
272  // happen with loops.
273  if (MI->getParent() == SuccToSinkTo)
274    return false;
275
276  DEBUG(dbgs() << "Sink instr " << *MI);
277  DEBUG(dbgs() << "to block " << *SuccToSinkTo);
278
279  // If the block has multiple predecessors, this would introduce computation on
280  // a path that it doesn't already exist.  We could split the critical edge,
281  // but for now we just punt.
282  // FIXME: Split critical edges if not backedges.
283  if (SuccToSinkTo->pred_size() > 1) {
284    // We cannot sink a load across a critical edge - there may be stores in
285    // other code paths.
286    bool store = true;
287    if (!MI->isSafeToMove(TII, AA, store)) {
288      DEBUG(dbgs() << " *** PUNTING: Wont sink load along critical edge.\n");
289      return false;
290    }
291
292    // We don't want to sink across a critical edge if we don't dominate the
293    // successor. We could be introducing calculations to new code paths.
294    if (!DT->dominates(ParentBlock, SuccToSinkTo)) {
295      DEBUG(dbgs() << " *** PUNTING: Critical edge found\n");
296      return false;
297    }
298
299    // Don't sink instructions into a loop.
300    if (LI->isLoopHeader(SuccToSinkTo)) {
301      DEBUG(dbgs() << " *** PUNTING: Loop header found\n");
302      return false;
303    }
304
305    // Otherwise we are OK with sinking along a critical edge.
306    DEBUG(dbgs() << "Sinking along critical edge.\n");
307  }
308
309  // Determine where to insert into.  Skip phi nodes.
310  MachineBasicBlock::iterator InsertPos = SuccToSinkTo->begin();
311  while (InsertPos != SuccToSinkTo->end() && InsertPos->isPHI())
312    ++InsertPos;
313
314  // Move the instruction.
315  SuccToSinkTo->splice(InsertPos, ParentBlock, MI,
316                       ++MachineBasicBlock::iterator(MI));
317
318  // Conservatively, clear any kill flags, since it's possible that
319  // they are no longer correct.
320  MI->clearKillInfo();
321
322  return true;
323}
324