ArgumentPromotion.cpp revision 28252b6f0a483ffb0ead991c7a1ead14e3cd2fc1
1//===-- ArgumentPromotion.cpp - Promote by-reference arguments ------------===//
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 promotes "by reference" arguments to be "by value" arguments.  In
11// practice, this means looking for internal functions that have pointer
12// arguments.  If it can prove, through the use of alias analysis, that an
13// argument is *only* loaded, then it can pass the value into the function
14// instead of the address of the value.  This can cause recursive simplification
15// of code and lead to the elimination of allocas (especially in C++ template
16// code like the STL).
17//
18// This pass also handles aggregate arguments that are passed into a function,
19// scalarizing them if the elements of the aggregate are only loaded.  Note that
20// by default it refuses to scalarize aggregates which would require passing in
21// more than three operands to the function, because passing thousands of
22// operands for a large array or structure is unprofitable! This limit can be
23// configured or disabled, however.
24//
25// Note that this transformation could also be done for arguments that are only
26// stored to (returning the value instead), but does not currently.  This case
27// would be best handled when and if LLVM begins supporting multiple return
28// values from functions.
29//
30//===----------------------------------------------------------------------===//
31
32#define DEBUG_TYPE "argpromotion"
33#include "llvm/Transforms/IPO.h"
34#include "llvm/Constants.h"
35#include "llvm/DerivedTypes.h"
36#include "llvm/Module.h"
37#include "llvm/CallGraphSCCPass.h"
38#include "llvm/Instructions.h"
39#include "llvm/LLVMContext.h"
40#include "llvm/Analysis/AliasAnalysis.h"
41#include "llvm/Analysis/CallGraph.h"
42#include "llvm/Support/CallSite.h"
43#include "llvm/Support/CFG.h"
44#include "llvm/Support/Debug.h"
45#include "llvm/Support/raw_ostream.h"
46#include "llvm/ADT/DepthFirstIterator.h"
47#include "llvm/ADT/Statistic.h"
48#include "llvm/ADT/StringExtras.h"
49#include <set>
50using namespace llvm;
51
52STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted");
53STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted");
54STATISTIC(NumByValArgsPromoted , "Number of byval arguments promoted");
55STATISTIC(NumArgumentsDead     , "Number of dead pointer args eliminated");
56
57namespace {
58  /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass.
59  ///
60  struct ArgPromotion : public CallGraphSCCPass {
61    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
62      AU.addRequired<AliasAnalysis>();
63      CallGraphSCCPass::getAnalysisUsage(AU);
64    }
65
66    virtual bool runOnSCC(CallGraphSCC &SCC);
67    static char ID; // Pass identification, replacement for typeid
68    explicit ArgPromotion(unsigned maxElements = 3)
69        : CallGraphSCCPass(ID), maxElements(maxElements) {
70      initializeArgPromotionPass(*PassRegistry::getPassRegistry());
71    }
72
73    /// A vector used to hold the indices of a single GEP instruction
74    typedef std::vector<uint64_t> IndicesVector;
75
76  private:
77    CallGraphNode *PromoteArguments(CallGraphNode *CGN);
78    bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const;
79    CallGraphNode *DoPromotion(Function *F,
80                               SmallPtrSet<Argument*, 8> &ArgsToPromote,
81                               SmallPtrSet<Argument*, 8> &ByValArgsToTransform);
82    /// The maximum number of elements to expand, or 0 for unlimited.
83    unsigned maxElements;
84  };
85}
86
87char ArgPromotion::ID = 0;
88INITIALIZE_PASS_BEGIN(ArgPromotion, "argpromotion",
89                "Promote 'by reference' arguments to scalars", false, false)
90INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
91INITIALIZE_AG_DEPENDENCY(CallGraph)
92INITIALIZE_PASS_END(ArgPromotion, "argpromotion",
93                "Promote 'by reference' arguments to scalars", false, false)
94
95Pass *llvm::createArgumentPromotionPass(unsigned maxElements) {
96  return new ArgPromotion(maxElements);
97}
98
99bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) {
100  bool Changed = false, LocalChange;
101
102  do {  // Iterate until we stop promoting from this SCC.
103    LocalChange = false;
104    // Attempt to promote arguments from all functions in this SCC.
105    for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
106      if (CallGraphNode *CGN = PromoteArguments(*I)) {
107        LocalChange = true;
108        SCC.ReplaceNode(*I, CGN);
109      }
110    }
111    Changed |= LocalChange;               // Remember that we changed something.
112  } while (LocalChange);
113
114  return Changed;
115}
116
117/// PromoteArguments - This method checks the specified function to see if there
118/// are any promotable arguments and if it is safe to promote the function (for
119/// example, all callers are direct).  If safe to promote some arguments, it
120/// calls the DoPromotion method.
121///
122CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) {
123  Function *F = CGN->getFunction();
124
125  // Make sure that it is local to this module.
126  if (!F || !F->hasLocalLinkage()) return 0;
127
128  // First check: see if there are any pointer arguments!  If not, quick exit.
129  SmallVector<std::pair<Argument*, unsigned>, 16> PointerArgs;
130  unsigned ArgNo = 0;
131  for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
132       I != E; ++I, ++ArgNo)
133    if (I->getType()->isPointerTy())
134      PointerArgs.push_back(std::pair<Argument*, unsigned>(I, ArgNo));
135  if (PointerArgs.empty()) return 0;
136
137  // Second check: make sure that all callers are direct callers.  We can't
138  // transform functions that have indirect callers.  Also see if the function
139  // is self-recursive.
140  bool isSelfRecursive = false;
141  for (Value::use_iterator UI = F->use_begin(), E = F->use_end();
142       UI != E; ++UI) {
143    CallSite CS(*UI);
144    // Must be a direct call.
145    if (CS.getInstruction() == 0 || !CS.isCallee(UI)) return 0;
146
147    if (CS.getInstruction()->getParent()->getParent() == F)
148      isSelfRecursive = true;
149  }
150
151  // Check to see which arguments are promotable.  If an argument is promotable,
152  // add it to ArgsToPromote.
153  SmallPtrSet<Argument*, 8> ArgsToPromote;
154  SmallPtrSet<Argument*, 8> ByValArgsToTransform;
155  for (unsigned i = 0; i != PointerArgs.size(); ++i) {
156    bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal);
157    Argument *PtrArg = PointerArgs[i].first;
158    const Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType();
159
160    // If this is a byval argument, and if the aggregate type is small, just
161    // pass the elements, which is always safe.
162    if (isByVal) {
163      if (const StructType *STy = dyn_cast<StructType>(AgTy)) {
164        if (maxElements > 0 && STy->getNumElements() > maxElements) {
165          DEBUG(dbgs() << "argpromotion disable promoting argument '"
166                << PtrArg->getName() << "' because it would require adding more"
167                << " than " << maxElements << " arguments to the function.\n");
168          continue;
169        }
170
171        // If all the elements are single-value types, we can promote it.
172        bool AllSimple = true;
173        for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
174          if (!STy->getElementType(i)->isSingleValueType()) {
175            AllSimple = false;
176            break;
177          }
178        }
179
180        // Safe to transform, don't even bother trying to "promote" it.
181        // Passing the elements as a scalar will allow scalarrepl to hack on
182        // the new alloca we introduce.
183        if (AllSimple) {
184          ByValArgsToTransform.insert(PtrArg);
185          continue;
186        }
187      }
188    }
189
190    // If the argument is a recursive type and we're in a recursive
191    // function, we could end up infinitely peeling the function argument.
192    if (isSelfRecursive) {
193      if (const StructType *STy = dyn_cast<StructType>(AgTy)) {
194        bool RecursiveType = false;
195        for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
196          if (STy->getElementType(i) == PtrArg->getType()) {
197            RecursiveType = true;
198            break;
199          }
200        }
201        if (RecursiveType)
202          continue;
203      }
204    }
205
206    // Otherwise, see if we can promote the pointer to its value.
207    if (isSafeToPromoteArgument(PtrArg, isByVal))
208      ArgsToPromote.insert(PtrArg);
209  }
210
211  // No promotable pointer arguments.
212  if (ArgsToPromote.empty() && ByValArgsToTransform.empty())
213    return 0;
214
215  return DoPromotion(F, ArgsToPromote, ByValArgsToTransform);
216}
217
218/// AllCallersPassInValidPointerForArgument - Return true if we can prove that
219/// all callees pass in a valid pointer for the specified function argument.
220static bool AllCallersPassInValidPointerForArgument(Argument *Arg) {
221  Function *Callee = Arg->getParent();
222
223  unsigned ArgNo = std::distance(Callee->arg_begin(),
224                                 Function::arg_iterator(Arg));
225
226  // Look at all call sites of the function.  At this pointer we know we only
227  // have direct callees.
228  for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end();
229       UI != E; ++UI) {
230    CallSite CS(*UI);
231    assert(CS && "Should only have direct calls!");
232
233    if (!CS.getArgument(ArgNo)->isDereferenceablePointer())
234      return false;
235  }
236  return true;
237}
238
239/// Returns true if Prefix is a prefix of longer. That means, Longer has a size
240/// that is greater than or equal to the size of prefix, and each of the
241/// elements in Prefix is the same as the corresponding elements in Longer.
242///
243/// This means it also returns true when Prefix and Longer are equal!
244static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix,
245                     const ArgPromotion::IndicesVector &Longer) {
246  if (Prefix.size() > Longer.size())
247    return false;
248  for (unsigned i = 0, e = Prefix.size(); i != e; ++i)
249    if (Prefix[i] != Longer[i])
250      return false;
251  return true;
252}
253
254
255/// Checks if Indices, or a prefix of Indices, is in Set.
256static bool PrefixIn(const ArgPromotion::IndicesVector &Indices,
257                     std::set<ArgPromotion::IndicesVector> &Set) {
258    std::set<ArgPromotion::IndicesVector>::iterator Low;
259    Low = Set.upper_bound(Indices);
260    if (Low != Set.begin())
261      Low--;
262    // Low is now the last element smaller than or equal to Indices. This means
263    // it points to a prefix of Indices (possibly Indices itself), if such
264    // prefix exists.
265    //
266    // This load is safe if any prefix of its operands is safe to load.
267    return Low != Set.end() && IsPrefix(*Low, Indices);
268}
269
270/// Mark the given indices (ToMark) as safe in the given set of indices
271/// (Safe). Marking safe usually means adding ToMark to Safe. However, if there
272/// is already a prefix of Indices in Safe, Indices are implicitely marked safe
273/// already. Furthermore, any indices that Indices is itself a prefix of, are
274/// removed from Safe (since they are implicitely safe because of Indices now).
275static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark,
276                            std::set<ArgPromotion::IndicesVector> &Safe) {
277  std::set<ArgPromotion::IndicesVector>::iterator Low;
278  Low = Safe.upper_bound(ToMark);
279  // Guard against the case where Safe is empty
280  if (Low != Safe.begin())
281    Low--;
282  // Low is now the last element smaller than or equal to Indices. This
283  // means it points to a prefix of Indices (possibly Indices itself), if
284  // such prefix exists.
285  if (Low != Safe.end()) {
286    if (IsPrefix(*Low, ToMark))
287      // If there is already a prefix of these indices (or exactly these
288      // indices) marked a safe, don't bother adding these indices
289      return;
290
291    // Increment Low, so we can use it as a "insert before" hint
292    ++Low;
293  }
294  // Insert
295  Low = Safe.insert(Low, ToMark);
296  ++Low;
297  // If there we're a prefix of longer index list(s), remove those
298  std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end();
299  while (Low != End && IsPrefix(ToMark, *Low)) {
300    std::set<ArgPromotion::IndicesVector>::iterator Remove = Low;
301    ++Low;
302    Safe.erase(Remove);
303  }
304}
305
306/// isSafeToPromoteArgument - As you might guess from the name of this method,
307/// it checks to see if it is both safe and useful to promote the argument.
308/// This method limits promotion of aggregates to only promote up to three
309/// elements of the aggregate in order to avoid exploding the number of
310/// arguments passed in.
311bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const {
312  typedef std::set<IndicesVector> GEPIndicesSet;
313
314  // Quick exit for unused arguments
315  if (Arg->use_empty())
316    return true;
317
318  // We can only promote this argument if all of the uses are loads, or are GEP
319  // instructions (with constant indices) that are subsequently loaded.
320  //
321  // Promoting the argument causes it to be loaded in the caller
322  // unconditionally. This is only safe if we can prove that either the load
323  // would have happened in the callee anyway (ie, there is a load in the entry
324  // block) or the pointer passed in at every call site is guaranteed to be
325  // valid.
326  // In the former case, invalid loads can happen, but would have happened
327  // anyway, in the latter case, invalid loads won't happen. This prevents us
328  // from introducing an invalid load that wouldn't have happened in the
329  // original code.
330  //
331  // This set will contain all sets of indices that are loaded in the entry
332  // block, and thus are safe to unconditionally load in the caller.
333  GEPIndicesSet SafeToUnconditionallyLoad;
334
335  // This set contains all the sets of indices that we are planning to promote.
336  // This makes it possible to limit the number of arguments added.
337  GEPIndicesSet ToPromote;
338
339  // If the pointer is always valid, any load with first index 0 is valid.
340  if (isByVal || AllCallersPassInValidPointerForArgument(Arg))
341    SafeToUnconditionallyLoad.insert(IndicesVector(1, 0));
342
343  // First, iterate the entry block and mark loads of (geps of) arguments as
344  // safe.
345  BasicBlock *EntryBlock = Arg->getParent()->begin();
346  // Declare this here so we can reuse it
347  IndicesVector Indices;
348  for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end();
349       I != E; ++I)
350    if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
351      Value *V = LI->getPointerOperand();
352      if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
353        V = GEP->getPointerOperand();
354        if (V == Arg) {
355          // This load actually loads (part of) Arg? Check the indices then.
356          Indices.reserve(GEP->getNumIndices());
357          for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
358               II != IE; ++II)
359            if (ConstantInt *CI = dyn_cast<ConstantInt>(*II))
360              Indices.push_back(CI->getSExtValue());
361            else
362              // We found a non-constant GEP index for this argument? Bail out
363              // right away, can't promote this argument at all.
364              return false;
365
366          // Indices checked out, mark them as safe
367          MarkIndicesSafe(Indices, SafeToUnconditionallyLoad);
368          Indices.clear();
369        }
370      } else if (V == Arg) {
371        // Direct loads are equivalent to a GEP with a single 0 index.
372        MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad);
373      }
374    }
375
376  // Now, iterate all uses of the argument to see if there are any uses that are
377  // not (GEP+)loads, or any (GEP+)loads that are not safe to promote.
378  SmallVector<LoadInst*, 16> Loads;
379  IndicesVector Operands;
380  for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end();
381       UI != E; ++UI) {
382    User *U = *UI;
383    Operands.clear();
384    if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
385      if (LI->isVolatile()) return false;  // Don't hack volatile loads
386      Loads.push_back(LI);
387      // Direct loads are equivalent to a GEP with a zero index and then a load.
388      Operands.push_back(0);
389    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
390      if (GEP->use_empty()) {
391        // Dead GEP's cause trouble later.  Just remove them if we run into
392        // them.
393        getAnalysis<AliasAnalysis>().deleteValue(GEP);
394        GEP->eraseFromParent();
395        // TODO: This runs the above loop over and over again for dead GEPs
396        // Couldn't we just do increment the UI iterator earlier and erase the
397        // use?
398        return isSafeToPromoteArgument(Arg, isByVal);
399      }
400
401      // Ensure that all of the indices are constants.
402      for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end();
403        i != e; ++i)
404        if (ConstantInt *C = dyn_cast<ConstantInt>(*i))
405          Operands.push_back(C->getSExtValue());
406        else
407          return false;  // Not a constant operand GEP!
408
409      // Ensure that the only users of the GEP are load instructions.
410      for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end();
411           UI != E; ++UI)
412        if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
413          if (LI->isVolatile()) return false;  // Don't hack volatile loads
414          Loads.push_back(LI);
415        } else {
416          // Other uses than load?
417          return false;
418        }
419    } else {
420      return false;  // Not a load or a GEP.
421    }
422
423    // Now, see if it is safe to promote this load / loads of this GEP. Loading
424    // is safe if Operands, or a prefix of Operands, is marked as safe.
425    if (!PrefixIn(Operands, SafeToUnconditionallyLoad))
426      return false;
427
428    // See if we are already promoting a load with these indices. If not, check
429    // to make sure that we aren't promoting too many elements.  If so, nothing
430    // to do.
431    if (ToPromote.find(Operands) == ToPromote.end()) {
432      if (maxElements > 0 && ToPromote.size() == maxElements) {
433        DEBUG(dbgs() << "argpromotion not promoting argument '"
434              << Arg->getName() << "' because it would require adding more "
435              << "than " << maxElements << " arguments to the function.\n");
436        // We limit aggregate promotion to only promoting up to a fixed number
437        // of elements of the aggregate.
438        return false;
439      }
440      ToPromote.insert(Operands);
441    }
442  }
443
444  if (Loads.empty()) return true;  // No users, this is a dead argument.
445
446  // Okay, now we know that the argument is only used by load instructions and
447  // it is safe to unconditionally perform all of them. Use alias analysis to
448  // check to see if the pointer is guaranteed to not be modified from entry of
449  // the function to each of the load instructions.
450
451  // Because there could be several/many load instructions, remember which
452  // blocks we know to be transparent to the load.
453  SmallPtrSet<BasicBlock*, 16> TranspBlocks;
454
455  AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
456
457  for (unsigned i = 0, e = Loads.size(); i != e; ++i) {
458    // Check to see if the load is invalidated from the start of the block to
459    // the load itself.
460    LoadInst *Load = Loads[i];
461    BasicBlock *BB = Load->getParent();
462
463    AliasAnalysis::Location Loc = AA.getLocation(Load);
464    if (AA.canInstructionRangeModify(BB->front(), *Load, Loc))
465      return false;  // Pointer is invalidated!
466
467    // Now check every path from the entry block to the load for transparency.
468    // To do this, we perform a depth first search on the inverse CFG from the
469    // loading block.
470    for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
471      BasicBlock *P = *PI;
472      for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> >
473             I = idf_ext_begin(P, TranspBlocks),
474             E = idf_ext_end(P, TranspBlocks); I != E; ++I)
475        if (AA.canBasicBlockModify(**I, Loc))
476          return false;
477    }
478  }
479
480  // If the path from the entry of the function to each load is free of
481  // instructions that potentially invalidate the load, we can make the
482  // transformation!
483  return true;
484}
485
486/// DoPromotion - This method actually performs the promotion of the specified
487/// arguments, and returns the new function.  At this point, we know that it's
488/// safe to do so.
489CallGraphNode *ArgPromotion::DoPromotion(Function *F,
490                               SmallPtrSet<Argument*, 8> &ArgsToPromote,
491                              SmallPtrSet<Argument*, 8> &ByValArgsToTransform) {
492
493  // Start by computing a new prototype for the function, which is the same as
494  // the old function, but has modified arguments.
495  const FunctionType *FTy = F->getFunctionType();
496  std::vector<const Type*> Params;
497
498  typedef std::set<IndicesVector> ScalarizeTable;
499
500  // ScalarizedElements - If we are promoting a pointer that has elements
501  // accessed out of it, keep track of which elements are accessed so that we
502  // can add one argument for each.
503  //
504  // Arguments that are directly loaded will have a zero element value here, to
505  // handle cases where there are both a direct load and GEP accesses.
506  //
507  std::map<Argument*, ScalarizeTable> ScalarizedElements;
508
509  // OriginalLoads - Keep track of a representative load instruction from the
510  // original function so that we can tell the alias analysis implementation
511  // what the new GEP/Load instructions we are inserting look like.
512  std::map<IndicesVector, LoadInst*> OriginalLoads;
513
514  // Attributes - Keep track of the parameter attributes for the arguments
515  // that we are *not* promoting. For the ones that we do promote, the parameter
516  // attributes are lost
517  SmallVector<AttributeWithIndex, 8> AttributesVec;
518  const AttrListPtr &PAL = F->getAttributes();
519
520  // Add any return attributes.
521  if (Attributes attrs = PAL.getRetAttributes())
522    AttributesVec.push_back(AttributeWithIndex::get(0, attrs));
523
524  // First, determine the new argument list
525  unsigned ArgIndex = 1;
526  for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
527       ++I, ++ArgIndex) {
528    if (ByValArgsToTransform.count(I)) {
529      // Simple byval argument? Just add all the struct element types.
530      const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
531      const StructType *STy = cast<StructType>(AgTy);
532      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
533        Params.push_back(STy->getElementType(i));
534      ++NumByValArgsPromoted;
535    } else if (!ArgsToPromote.count(I)) {
536      // Unchanged argument
537      Params.push_back(I->getType());
538      if (Attributes attrs = PAL.getParamAttributes(ArgIndex))
539        AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs));
540    } else if (I->use_empty()) {
541      // Dead argument (which are always marked as promotable)
542      ++NumArgumentsDead;
543    } else {
544      // Okay, this is being promoted. This means that the only uses are loads
545      // or GEPs which are only used by loads
546
547      // In this table, we will track which indices are loaded from the argument
548      // (where direct loads are tracked as no indices).
549      ScalarizeTable &ArgIndices = ScalarizedElements[I];
550      for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
551           ++UI) {
552        Instruction *User = cast<Instruction>(*UI);
553        assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User));
554        IndicesVector Indices;
555        Indices.reserve(User->getNumOperands() - 1);
556        // Since loads will only have a single operand, and GEPs only a single
557        // non-index operand, this will record direct loads without any indices,
558        // and gep+loads with the GEP indices.
559        for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end();
560             II != IE; ++II)
561          Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
562        // GEPs with a single 0 index can be merged with direct loads
563        if (Indices.size() == 1 && Indices.front() == 0)
564          Indices.clear();
565        ArgIndices.insert(Indices);
566        LoadInst *OrigLoad;
567        if (LoadInst *L = dyn_cast<LoadInst>(User))
568          OrigLoad = L;
569        else
570          // Take any load, we will use it only to update Alias Analysis
571          OrigLoad = cast<LoadInst>(User->use_back());
572        OriginalLoads[Indices] = OrigLoad;
573      }
574
575      // Add a parameter to the function for each element passed in.
576      for (ScalarizeTable::iterator SI = ArgIndices.begin(),
577             E = ArgIndices.end(); SI != E; ++SI) {
578        // not allowed to dereference ->begin() if size() is 0
579        Params.push_back(GetElementPtrInst::getIndexedType(I->getType(),
580                                                           SI->begin(),
581                                                           SI->end()));
582        assert(Params.back());
583      }
584
585      if (ArgIndices.size() == 1 && ArgIndices.begin()->empty())
586        ++NumArgumentsPromoted;
587      else
588        ++NumAggregatesPromoted;
589    }
590  }
591
592  // Add any function attributes.
593  if (Attributes attrs = PAL.getFnAttributes())
594    AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));
595
596  const Type *RetTy = FTy->getReturnType();
597
598  // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which
599  // have zero fixed arguments.
600  bool ExtraArgHack = false;
601  if (Params.empty() && FTy->isVarArg()) {
602    ExtraArgHack = true;
603    Params.push_back(Type::getInt32Ty(F->getContext()));
604  }
605
606  // Construct the new function type using the new arguments.
607  FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
608
609  // Create the new function body and insert it into the module.
610  Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
611  NF->copyAttributesFrom(F);
612
613
614  DEBUG(dbgs() << "ARG PROMOTION:  Promoting to:" << *NF << "\n"
615        << "From: " << *F);
616
617  // Recompute the parameter attributes list based on the new arguments for
618  // the function.
619  NF->setAttributes(AttrListPtr::get(AttributesVec.begin(),
620                                     AttributesVec.end()));
621  AttributesVec.clear();
622
623  F->getParent()->getFunctionList().insert(F, NF);
624  NF->takeName(F);
625
626  // Get the alias analysis information that we need to update to reflect our
627  // changes.
628  AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
629
630  // Get the callgraph information that we need to update to reflect our
631  // changes.
632  CallGraph &CG = getAnalysis<CallGraph>();
633
634  // Get a new callgraph node for NF.
635  CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
636
637  // Loop over all of the callers of the function, transforming the call sites
638  // to pass in the loaded pointers.
639  //
640  SmallVector<Value*, 16> Args;
641  while (!F->use_empty()) {
642    CallSite CS(F->use_back());
643    assert(CS.getCalledFunction() == F);
644    Instruction *Call = CS.getInstruction();
645    const AttrListPtr &CallPAL = CS.getAttributes();
646
647    // Add any return attributes.
648    if (Attributes attrs = CallPAL.getRetAttributes())
649      AttributesVec.push_back(AttributeWithIndex::get(0, attrs));
650
651    // Loop over the operands, inserting GEP and loads in the caller as
652    // appropriate.
653    CallSite::arg_iterator AI = CS.arg_begin();
654    ArgIndex = 1;
655    for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
656         I != E; ++I, ++AI, ++ArgIndex)
657      if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
658        Args.push_back(*AI);          // Unmodified argument
659
660        if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
661          AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
662
663      } else if (ByValArgsToTransform.count(I)) {
664        // Emit a GEP and load for each element of the struct.
665        const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
666        const StructType *STy = cast<StructType>(AgTy);
667        Value *Idxs[2] = {
668              ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
669        for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
670          Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
671          Value *Idx = GetElementPtrInst::Create(*AI, Idxs, Idxs+2,
672                                                 (*AI)->getName()+"."+utostr(i),
673                                                 Call);
674          // TODO: Tell AA about the new values?
675          Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
676        }
677      } else if (!I->use_empty()) {
678        // Non-dead argument: insert GEPs and loads as appropriate.
679        ScalarizeTable &ArgIndices = ScalarizedElements[I];
680        // Store the Value* version of the indices in here, but declare it now
681        // for reuse.
682        std::vector<Value*> Ops;
683        for (ScalarizeTable::iterator SI = ArgIndices.begin(),
684               E = ArgIndices.end(); SI != E; ++SI) {
685          Value *V = *AI;
686          LoadInst *OrigLoad = OriginalLoads[*SI];
687          if (!SI->empty()) {
688            Ops.reserve(SI->size());
689            const Type *ElTy = V->getType();
690            for (IndicesVector::const_iterator II = SI->begin(),
691                 IE = SI->end(); II != IE; ++II) {
692              // Use i32 to index structs, and i64 for others (pointers/arrays).
693              // This satisfies GEP constraints.
694              const Type *IdxTy = (ElTy->isStructTy() ?
695                    Type::getInt32Ty(F->getContext()) :
696                    Type::getInt64Ty(F->getContext()));
697              Ops.push_back(ConstantInt::get(IdxTy, *II));
698              // Keep track of the type we're currently indexing.
699              ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
700            }
701            // And create a GEP to extract those indices.
702            V = GetElementPtrInst::Create(V, Ops.begin(), Ops.end(),
703                                          V->getName()+".idx", Call);
704            Ops.clear();
705            AA.copyValue(OrigLoad->getOperand(0), V);
706          }
707          // Since we're replacing a load make sure we take the alignment
708          // of the previous load.
709          LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
710          newLoad->setAlignment(OrigLoad->getAlignment());
711          // Transfer the TBAA info too.
712          newLoad->setMetadata(LLVMContext::MD_tbaa,
713                               OrigLoad->getMetadata(LLVMContext::MD_tbaa));
714          Args.push_back(newLoad);
715          AA.copyValue(OrigLoad, Args.back());
716        }
717      }
718
719    if (ExtraArgHack)
720      Args.push_back(Constant::getNullValue(Type::getInt32Ty(F->getContext())));
721
722    // Push any varargs arguments on the list.
723    for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
724      Args.push_back(*AI);
725      if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
726        AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
727    }
728
729    // Add any function attributes.
730    if (Attributes attrs = CallPAL.getFnAttributes())
731      AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));
732
733    Instruction *New;
734    if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
735      New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
736                               Args.begin(), Args.end(), "", Call);
737      cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
738      cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
739                                                          AttributesVec.end()));
740    } else {
741      New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call);
742      cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
743      cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
744                                                        AttributesVec.end()));
745      if (cast<CallInst>(Call)->isTailCall())
746        cast<CallInst>(New)->setTailCall();
747    }
748    Args.clear();
749    AttributesVec.clear();
750
751    // Update the alias analysis implementation to know that we are replacing
752    // the old call with a new one.
753    AA.replaceWithNewValue(Call, New);
754
755    // Update the callgraph to know that the callsite has been transformed.
756    CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
757    CalleeNode->replaceCallEdge(Call, New, NF_CGN);
758
759    if (!Call->use_empty()) {
760      Call->replaceAllUsesWith(New);
761      New->takeName(Call);
762    }
763
764    // Finally, remove the old call from the program, reducing the use-count of
765    // F.
766    Call->eraseFromParent();
767  }
768
769  // Since we have now created the new function, splice the body of the old
770  // function right into the new function, leaving the old rotting hulk of the
771  // function empty.
772  NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
773
774  // Loop over the argument list, transfering uses of the old arguments over to
775  // the new arguments, also transfering over the names as well.
776  //
777  for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
778       I2 = NF->arg_begin(); I != E; ++I) {
779    if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
780      // If this is an unmodified argument, move the name and users over to the
781      // new version.
782      I->replaceAllUsesWith(I2);
783      I2->takeName(I);
784      AA.replaceWithNewValue(I, I2);
785      ++I2;
786      continue;
787    }
788
789    if (ByValArgsToTransform.count(I)) {
790      // In the callee, we create an alloca, and store each of the new incoming
791      // arguments into the alloca.
792      Instruction *InsertPt = NF->begin()->begin();
793
794      // Just add all the struct element types.
795      const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
796      Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt);
797      const StructType *STy = cast<StructType>(AgTy);
798      Value *Idxs[2] = {
799            ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
800
801      for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
802        Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
803        Value *Idx =
804          GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2,
805                                    TheAlloca->getName()+"."+Twine(i),
806                                    InsertPt);
807        I2->setName(I->getName()+"."+Twine(i));
808        new StoreInst(I2++, Idx, InsertPt);
809      }
810
811      // Anything that used the arg should now use the alloca.
812      I->replaceAllUsesWith(TheAlloca);
813      TheAlloca->takeName(I);
814      AA.replaceWithNewValue(I, TheAlloca);
815      continue;
816    }
817
818    if (I->use_empty()) {
819      AA.deleteValue(I);
820      continue;
821    }
822
823    // Otherwise, if we promoted this argument, then all users are load
824    // instructions (or GEPs with only load users), and all loads should be
825    // using the new argument that we added.
826    ScalarizeTable &ArgIndices = ScalarizedElements[I];
827
828    while (!I->use_empty()) {
829      if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) {
830        assert(ArgIndices.begin()->empty() &&
831               "Load element should sort to front!");
832        I2->setName(I->getName()+".val");
833        LI->replaceAllUsesWith(I2);
834        AA.replaceWithNewValue(LI, I2);
835        LI->eraseFromParent();
836        DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
837              << "' in function '" << F->getName() << "'\n");
838      } else {
839        GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back());
840        IndicesVector Operands;
841        Operands.reserve(GEP->getNumIndices());
842        for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
843             II != IE; ++II)
844          Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
845
846        // GEPs with a single 0 index can be merged with direct loads
847        if (Operands.size() == 1 && Operands.front() == 0)
848          Operands.clear();
849
850        Function::arg_iterator TheArg = I2;
851        for (ScalarizeTable::iterator It = ArgIndices.begin();
852             *It != Operands; ++It, ++TheArg) {
853          assert(It != ArgIndices.end() && "GEP not handled??");
854        }
855
856        std::string NewName = I->getName();
857        for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
858            NewName += "." + utostr(Operands[i]);
859        }
860        NewName += ".val";
861        TheArg->setName(NewName);
862
863        DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
864              << "' of function '" << NF->getName() << "'\n");
865
866        // All of the uses must be load instructions.  Replace them all with
867        // the argument specified by ArgNo.
868        while (!GEP->use_empty()) {
869          LoadInst *L = cast<LoadInst>(GEP->use_back());
870          L->replaceAllUsesWith(TheArg);
871          AA.replaceWithNewValue(L, TheArg);
872          L->eraseFromParent();
873        }
874        AA.deleteValue(GEP);
875        GEP->eraseFromParent();
876      }
877    }
878
879    // Increment I2 past all of the arguments added for this promoted pointer.
880    for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i)
881      ++I2;
882  }
883
884  // Notify the alias analysis implementation that we inserted a new argument.
885  if (ExtraArgHack)
886    AA.copyValue(Constant::getNullValue(Type::getInt32Ty(F->getContext())),
887                 NF->arg_begin());
888
889
890  // Tell the alias analysis that the old function is about to disappear.
891  AA.replaceWithNewValue(F, NF);
892
893
894  NF_CGN->stealCalledFunctionsFrom(CG[F]);
895
896  // Now that the old function is dead, delete it.  If there is a dangling
897  // reference to the CallgraphNode, just leave the dead function around for
898  // someone else to nuke.
899  CallGraphNode *CGN = CG[F];
900  if (CGN->getNumReferences() == 0)
901    delete CG.removeFunctionFromModule(CGN);
902  else
903    F->setLinkage(Function::ExternalLinkage);
904
905  return NF_CGN;
906}
907