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