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/CallGraphSCCPass.h"
40 #include "llvm/Constants.h"
41 #include "llvm/DerivedTypes.h"
42 #include "llvm/Instructions.h"
43 #include "llvm/LLVMContext.h"
44 #include "llvm/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_AG_DEPENDENCY(CallGraph)
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<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;
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;
147 if (CS.getInstruction()->getParent()->getParent() == F)
148 isSelfRecursive = true;
149 }
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->getAttributes().
157 hasAttribute(PointerArgs[i].second+1, Attribute::ByVal);
158 Argument *PtrArg = PointerArgs[i].first;
159 Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType();
161 // If this is a byval argument, and if the aggregate type is small, just
162 // pass the elements, which is always safe.
163 if (isByVal) {
164 if (StructType *STy = dyn_cast<StructType>(AgTy)) {
165 if (maxElements > 0 && STy->getNumElements() > maxElements) {
166 DEBUG(dbgs() << "argpromotion disable promoting argument '"
167 << PtrArg->getName() << "' because it would require adding more"
168 << " than " << maxElements << " arguments to the function.\n");
169 continue;
170 }
172 // If all the elements are single-value types, we can promote it.
173 bool AllSimple = true;
174 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
175 if (!STy->getElementType(i)->isSingleValueType()) {
176 AllSimple = false;
177 break;
178 }
179 }
181 // Safe to transform, don't even bother trying to "promote" it.
182 // Passing the elements as a scalar will allow scalarrepl to hack on
183 // the new alloca we introduce.
184 if (AllSimple) {
185 ByValArgsToTransform.insert(PtrArg);
186 continue;
187 }
188 }
189 }
191 // If the argument is a recursive type and we're in a recursive
192 // function, we could end up infinitely peeling the function argument.
193 if (isSelfRecursive) {
194 if (StructType *STy = dyn_cast<StructType>(AgTy)) {
195 bool RecursiveType = false;
196 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
197 if (STy->getElementType(i) == PtrArg->getType()) {
198 RecursiveType = true;
199 break;
200 }
201 }
202 if (RecursiveType)
203 continue;
204 }
205 }
207 // Otherwise, see if we can promote the pointer to its value.
208 if (isSafeToPromoteArgument(PtrArg, isByVal))
209 ArgsToPromote.insert(PtrArg);
210 }
212 // No promotable pointer arguments.
213 if (ArgsToPromote.empty() && ByValArgsToTransform.empty())
214 return 0;
216 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform);
217 }
219 /// AllCallersPassInValidPointerForArgument - Return true if we can prove that
220 /// all callees pass in a valid pointer for the specified function argument.
221 static bool AllCallersPassInValidPointerForArgument(Argument *Arg) {
222 Function *Callee = Arg->getParent();
224 unsigned ArgNo = std::distance(Callee->arg_begin(),
225 Function::arg_iterator(Arg));
227 // Look at all call sites of the function. At this pointer we know we only
228 // have direct callees.
229 for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end();
230 UI != E; ++UI) {
231 CallSite CS(*UI);
232 assert(CS && "Should only have direct calls!");
234 if (!CS.getArgument(ArgNo)->isDereferenceablePointer())
235 return false;
236 }
237 return true;
238 }
240 /// Returns true if Prefix is a prefix of longer. That means, Longer has a size
241 /// that is greater than or equal to the size of prefix, and each of the
242 /// elements in Prefix is the same as the corresponding elements in Longer.
243 ///
244 /// This means it also returns true when Prefix and Longer are equal!
245 static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix,
246 const ArgPromotion::IndicesVector &Longer) {
247 if (Prefix.size() > Longer.size())
248 return false;
249 return std::equal(Prefix.begin(), Prefix.end(), Longer.begin());
250 }
253 /// Checks if Indices, or a prefix of Indices, is in Set.
254 static bool PrefixIn(const ArgPromotion::IndicesVector &Indices,
255 std::set<ArgPromotion::IndicesVector> &Set) {
256 std::set<ArgPromotion::IndicesVector>::iterator Low;
257 Low = Set.upper_bound(Indices);
258 if (Low != Set.begin())
259 Low--;
260 // Low is now the last element smaller than or equal to Indices. This means
261 // it points to a prefix of Indices (possibly Indices itself), if such
262 // prefix exists.
263 //
264 // This load is safe if any prefix of its operands is safe to load.
265 return Low != Set.end() && IsPrefix(*Low, Indices);
266 }
268 /// Mark the given indices (ToMark) as safe in the given set of indices
269 /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there
270 /// is already a prefix of Indices in Safe, Indices are implicitely marked safe
271 /// already. Furthermore, any indices that Indices is itself a prefix of, are
272 /// removed from Safe (since they are implicitely safe because of Indices now).
273 static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark,
274 std::set<ArgPromotion::IndicesVector> &Safe) {
275 std::set<ArgPromotion::IndicesVector>::iterator Low;
276 Low = Safe.upper_bound(ToMark);
277 // Guard against the case where Safe is empty
278 if (Low != Safe.begin())
279 Low--;
280 // Low is now the last element smaller than or equal to Indices. This
281 // means it points to a prefix of Indices (possibly Indices itself), if
282 // such prefix exists.
283 if (Low != Safe.end()) {
284 if (IsPrefix(*Low, ToMark))
285 // If there is already a prefix of these indices (or exactly these
286 // indices) marked a safe, don't bother adding these indices
287 return;
289 // Increment Low, so we can use it as a "insert before" hint
290 ++Low;
291 }
292 // Insert
293 Low = Safe.insert(Low, ToMark);
294 ++Low;
295 // If there we're a prefix of longer index list(s), remove those
296 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end();
297 while (Low != End && IsPrefix(ToMark, *Low)) {
298 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low;
299 ++Low;
300 Safe.erase(Remove);
301 }
302 }
304 /// isSafeToPromoteArgument - As you might guess from the name of this method,
305 /// it checks to see if it is both safe and useful to promote the argument.
306 /// This method limits promotion of aggregates to only promote up to three
307 /// elements of the aggregate in order to avoid exploding the number of
308 /// arguments passed in.
309 bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const {
310 typedef std::set<IndicesVector> GEPIndicesSet;
312 // Quick exit for unused arguments
313 if (Arg->use_empty())
314 return true;
316 // We can only promote this argument if all of the uses are loads, or are GEP
317 // instructions (with constant indices) that are subsequently loaded.
318 //
319 // Promoting the argument causes it to be loaded in the caller
320 // unconditionally. This is only safe if we can prove that either the load
321 // would have happened in the callee anyway (ie, there is a load in the entry
322 // block) or the pointer passed in at every call site is guaranteed to be
323 // valid.
324 // In the former case, invalid loads can happen, but would have happened
325 // anyway, in the latter case, invalid loads won't happen. This prevents us
326 // from introducing an invalid load that wouldn't have happened in the
327 // original code.
328 //
329 // This set will contain all sets of indices that are loaded in the entry
330 // block, and thus are safe to unconditionally load in the caller.
331 GEPIndicesSet SafeToUnconditionallyLoad;
333 // This set contains all the sets of indices that we are planning to promote.
334 // This makes it possible to limit the number of arguments added.
335 GEPIndicesSet ToPromote;
337 // If the pointer is always valid, any load with first index 0 is valid.
338 if (isByVal || AllCallersPassInValidPointerForArgument(Arg))
339 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0));
341 // First, iterate the entry block and mark loads of (geps of) arguments as
342 // safe.
343 BasicBlock *EntryBlock = Arg->getParent()->begin();
344 // Declare this here so we can reuse it
345 IndicesVector Indices;
346 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end();
347 I != E; ++I)
348 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
349 Value *V = LI->getPointerOperand();
350 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
351 V = GEP->getPointerOperand();
352 if (V == Arg) {
353 // This load actually loads (part of) Arg? Check the indices then.
354 Indices.reserve(GEP->getNumIndices());
355 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
356 II != IE; ++II)
357 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II))
358 Indices.push_back(CI->getSExtValue());
359 else
360 // We found a non-constant GEP index for this argument? Bail out
361 // right away, can't promote this argument at all.
362 return false;
364 // Indices checked out, mark them as safe
365 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad);
366 Indices.clear();
367 }
368 } else if (V == Arg) {
369 // Direct loads are equivalent to a GEP with a single 0 index.
370 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad);
371 }
372 }
374 // Now, iterate all uses of the argument to see if there are any uses that are
375 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote.
376 SmallVector<LoadInst*, 16> Loads;
377 IndicesVector Operands;
378 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end();
379 UI != E; ++UI) {
380 User *U = *UI;
381 Operands.clear();
382 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
383 // Don't hack volatile/atomic loads
384 if (!LI->isSimple()) return false;
385 Loads.push_back(LI);
386 // Direct loads are equivalent to a GEP with a zero index and then a load.
387 Operands.push_back(0);
388 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
389 if (GEP->use_empty()) {
390 // Dead GEP's cause trouble later. Just remove them if we run into
391 // them.
392 getAnalysis<AliasAnalysis>().deleteValue(GEP);
393 GEP->eraseFromParent();
394 // TODO: This runs the above loop over and over again for dead GEPs
395 // Couldn't we just do increment the UI iterator earlier and erase the
396 // use?
397 return isSafeToPromoteArgument(Arg, isByVal);
398 }
400 // Ensure that all of the indices are constants.
401 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end();
402 i != e; ++i)
403 if (ConstantInt *C = dyn_cast<ConstantInt>(*i))
404 Operands.push_back(C->getSExtValue());
405 else
406 return false; // Not a constant operand GEP!
408 // Ensure that the only users of the GEP are load instructions.
409 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end();
410 UI != E; ++UI)
411 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
412 // Don't hack volatile/atomic loads
413 if (!LI->isSimple()) return false;
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 }
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;
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 }
444 if (Loads.empty()) return true; // No users, this is a dead argument.
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.
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;
455 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
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();
463 AliasAnalysis::Location Loc = AA.getLocation(Load);
464 if (AA.canInstructionRangeModify(BB->front(), *Load, Loc))
465 return false; // Pointer is invalidated!
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 }
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 }
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.
489 CallGraphNode *ArgPromotion::DoPromotion(Function *F,
490 SmallPtrSet<Argument*, 8> &ArgsToPromote,
491 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) {
493 // Start by computing a new prototype for the function, which is the same as
494 // the old function, but has modified arguments.
495 FunctionType *FTy = F->getFunctionType();
496 std::vector<Type*> Params;
498 typedef std::set<IndicesVector> ScalarizeTable;
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;
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;
514 // Attribute - 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 AttributeSet &PAL = F->getAttributes();
520 // Add any return attributes.
521 Attribute attrs = PAL.getRetAttributes();
522 if (attrs.hasAttributes())
523 AttributesVec.push_back(AttributeWithIndex::get(AttributeSet::ReturnIndex,
524 attrs));
526 // First, determine the new argument list
527 unsigned ArgIndex = 1;
528 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
529 ++I, ++ArgIndex) {
530 if (ByValArgsToTransform.count(I)) {
531 // Simple byval argument? Just add all the struct element types.
532 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
533 StructType *STy = cast<StructType>(AgTy);
534 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
535 Params.push_back(STy->getElementType(i));
536 ++NumByValArgsPromoted;
537 } else if (!ArgsToPromote.count(I)) {
538 // Unchanged argument
539 Params.push_back(I->getType());
540 Attribute attrs = PAL.getParamAttributes(ArgIndex);
541 if (attrs.hasAttributes())
542 AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs));
543 } else if (I->use_empty()) {
544 // Dead argument (which are always marked as promotable)
545 ++NumArgumentsDead;
546 } else {
547 // Okay, this is being promoted. This means that the only uses are loads
548 // or GEPs which are only used by loads
550 // In this table, we will track which indices are loaded from the argument
551 // (where direct loads are tracked as no indices).
552 ScalarizeTable &ArgIndices = ScalarizedElements[I];
553 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
554 ++UI) {
555 Instruction *User = cast<Instruction>(*UI);
556 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User));
557 IndicesVector Indices;
558 Indices.reserve(User->getNumOperands() - 1);
559 // Since loads will only have a single operand, and GEPs only a single
560 // non-index operand, this will record direct loads without any indices,
561 // and gep+loads with the GEP indices.
562 for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end();
563 II != IE; ++II)
564 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
565 // GEPs with a single 0 index can be merged with direct loads
566 if (Indices.size() == 1 && Indices.front() == 0)
567 Indices.clear();
568 ArgIndices.insert(Indices);
569 LoadInst *OrigLoad;
570 if (LoadInst *L = dyn_cast<LoadInst>(User))
571 OrigLoad = L;
572 else
573 // Take any load, we will use it only to update Alias Analysis
574 OrigLoad = cast<LoadInst>(User->use_back());
575 OriginalLoads[Indices] = OrigLoad;
576 }
578 // Add a parameter to the function for each element passed in.
579 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
580 E = ArgIndices.end(); SI != E; ++SI) {
581 // not allowed to dereference ->begin() if size() is 0
582 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), *SI));
583 assert(Params.back());
584 }
586 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty())
587 ++NumArgumentsPromoted;
588 else
589 ++NumAggregatesPromoted;
590 }
591 }
593 // Add any function attributes.
594 attrs = PAL.getFnAttributes();
595 if (attrs.hasAttributes())
596 AttributesVec.push_back(AttributeWithIndex::get(AttributeSet::FunctionIndex,
597 attrs));
599 Type *RetTy = FTy->getReturnType();
601 // Construct the new function type using the new arguments.
602 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
604 // Create the new function body and insert it into the module.
605 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
606 NF->copyAttributesFrom(F);
609 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n"
610 << "From: " << *F);
612 // Recompute the parameter attributes list based on the new arguments for
613 // the function.
614 NF->setAttributes(AttributeSet::get(F->getContext(), AttributesVec));
615 AttributesVec.clear();
617 F->getParent()->getFunctionList().insert(F, NF);
618 NF->takeName(F);
620 // Get the alias analysis information that we need to update to reflect our
621 // changes.
622 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
624 // Get the callgraph information that we need to update to reflect our
625 // changes.
626 CallGraph &CG = getAnalysis<CallGraph>();
628 // Get a new callgraph node for NF.
629 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
631 // Loop over all of the callers of the function, transforming the call sites
632 // to pass in the loaded pointers.
633 //
634 SmallVector<Value*, 16> Args;
635 while (!F->use_empty()) {
636 CallSite CS(F->use_back());
637 assert(CS.getCalledFunction() == F);
638 Instruction *Call = CS.getInstruction();
639 const AttributeSet &CallPAL = CS.getAttributes();
641 // Add any return attributes.
642 Attribute attrs = CallPAL.getRetAttributes();
643 if (attrs.hasAttributes())
644 AttributesVec.push_back(AttributeWithIndex::get(AttributeSet::ReturnIndex,
645 attrs));
647 // Loop over the operands, inserting GEP and loads in the caller as
648 // appropriate.
649 CallSite::arg_iterator AI = CS.arg_begin();
650 ArgIndex = 1;
651 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
652 I != E; ++I, ++AI, ++ArgIndex)
653 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
654 Args.push_back(*AI); // Unmodified argument
656 Attribute Attrs = CallPAL.getParamAttributes(ArgIndex);
657 if (Attrs.hasAttributes())
658 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
660 } else if (ByValArgsToTransform.count(I)) {
661 // Emit a GEP and load for each element of the struct.
662 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
663 StructType *STy = cast<StructType>(AgTy);
664 Value *Idxs[2] = {
665 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
666 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
667 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
668 Value *Idx = GetElementPtrInst::Create(*AI, Idxs,
669 (*AI)->getName()+"."+utostr(i),
670 Call);
671 // TODO: Tell AA about the new values?
672 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
673 }
674 } else if (!I->use_empty()) {
675 // Non-dead argument: insert GEPs and loads as appropriate.
676 ScalarizeTable &ArgIndices = ScalarizedElements[I];
677 // Store the Value* version of the indices in here, but declare it now
678 // for reuse.
679 std::vector<Value*> Ops;
680 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
681 E = ArgIndices.end(); SI != E; ++SI) {
682 Value *V = *AI;
683 LoadInst *OrigLoad = OriginalLoads[*SI];
684 if (!SI->empty()) {
685 Ops.reserve(SI->size());
686 Type *ElTy = V->getType();
687 for (IndicesVector::const_iterator II = SI->begin(),
688 IE = SI->end(); II != IE; ++II) {
689 // Use i32 to index structs, and i64 for others (pointers/arrays).
690 // This satisfies GEP constraints.
691 Type *IdxTy = (ElTy->isStructTy() ?
692 Type::getInt32Ty(F->getContext()) :
693 Type::getInt64Ty(F->getContext()));
694 Ops.push_back(ConstantInt::get(IdxTy, *II));
695 // Keep track of the type we're currently indexing.
696 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
697 }
698 // And create a GEP to extract those indices.
699 V = GetElementPtrInst::Create(V, Ops, V->getName()+".idx", Call);
700 Ops.clear();
701 AA.copyValue(OrigLoad->getOperand(0), V);
702 }
703 // Since we're replacing a load make sure we take the alignment
704 // of the previous load.
705 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
706 newLoad->setAlignment(OrigLoad->getAlignment());
707 // Transfer the TBAA info too.
708 newLoad->setMetadata(LLVMContext::MD_tbaa,
709 OrigLoad->getMetadata(LLVMContext::MD_tbaa));
710 Args.push_back(newLoad);
711 AA.copyValue(OrigLoad, Args.back());
712 }
713 }
715 // Push any varargs arguments on the list.
716 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
717 Args.push_back(*AI);
718 Attribute Attrs = CallPAL.getParamAttributes(ArgIndex);
719 if (Attrs.hasAttributes())
720 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
721 }
723 // Add any function attributes.
724 attrs = CallPAL.getFnAttributes();
725 if (attrs.hasAttributes())
726 AttributesVec.push_back(AttributeWithIndex::get(AttributeSet::FunctionIndex,
727 attrs));
729 Instruction *New;
730 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
731 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
732 Args, "", Call);
733 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
734 cast<InvokeInst>(New)->setAttributes(AttributeSet::get(II->getContext(),
735 AttributesVec));
736 } else {
737 New = CallInst::Create(NF, Args, "", Call);
738 cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
739 cast<CallInst>(New)->setAttributes(AttributeSet::get(New->getContext(),
740 AttributesVec));
741 if (cast<CallInst>(Call)->isTailCall())
742 cast<CallInst>(New)->setTailCall();
743 }
744 Args.clear();
745 AttributesVec.clear();
747 // Update the alias analysis implementation to know that we are replacing
748 // the old call with a new one.
749 AA.replaceWithNewValue(Call, New);
751 // Update the callgraph to know that the callsite has been transformed.
752 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
753 CalleeNode->replaceCallEdge(Call, New, NF_CGN);
755 if (!Call->use_empty()) {
756 Call->replaceAllUsesWith(New);
757 New->takeName(Call);
758 }
760 // Finally, remove the old call from the program, reducing the use-count of
761 // F.
762 Call->eraseFromParent();
763 }
765 // Since we have now created the new function, splice the body of the old
766 // function right into the new function, leaving the old rotting hulk of the
767 // function empty.
768 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
770 // Loop over the argument list, transferring uses of the old arguments over to
771 // the new arguments, also transferring over the names as well.
772 //
773 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
774 I2 = NF->arg_begin(); I != E; ++I) {
775 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
776 // If this is an unmodified argument, move the name and users over to the
777 // new version.
778 I->replaceAllUsesWith(I2);
779 I2->takeName(I);
780 AA.replaceWithNewValue(I, I2);
781 ++I2;
782 continue;
783 }
785 if (ByValArgsToTransform.count(I)) {
786 // In the callee, we create an alloca, and store each of the new incoming
787 // arguments into the alloca.
788 Instruction *InsertPt = NF->begin()->begin();
790 // Just add all the struct element types.
791 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
792 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt);
793 StructType *STy = cast<StructType>(AgTy);
794 Value *Idxs[2] = {
795 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
797 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
798 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
799 Value *Idx =
800 GetElementPtrInst::Create(TheAlloca, Idxs,
801 TheAlloca->getName()+"."+Twine(i),
802 InsertPt);
803 I2->setName(I->getName()+"."+Twine(i));
804 new StoreInst(I2++, Idx, InsertPt);
805 }
807 // Anything that used the arg should now use the alloca.
808 I->replaceAllUsesWith(TheAlloca);
809 TheAlloca->takeName(I);
810 AA.replaceWithNewValue(I, TheAlloca);
811 continue;
812 }
814 if (I->use_empty()) {
815 AA.deleteValue(I);
816 continue;
817 }
819 // Otherwise, if we promoted this argument, then all users are load
820 // instructions (or GEPs with only load users), and all loads should be
821 // using the new argument that we added.
822 ScalarizeTable &ArgIndices = ScalarizedElements[I];
824 while (!I->use_empty()) {
825 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) {
826 assert(ArgIndices.begin()->empty() &&
827 "Load element should sort to front!");
828 I2->setName(I->getName()+".val");
829 LI->replaceAllUsesWith(I2);
830 AA.replaceWithNewValue(LI, I2);
831 LI->eraseFromParent();
832 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
833 << "' in function '" << F->getName() << "'\n");
834 } else {
835 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back());
836 IndicesVector Operands;
837 Operands.reserve(GEP->getNumIndices());
838 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
839 II != IE; ++II)
840 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
842 // GEPs with a single 0 index can be merged with direct loads
843 if (Operands.size() == 1 && Operands.front() == 0)
844 Operands.clear();
846 Function::arg_iterator TheArg = I2;
847 for (ScalarizeTable::iterator It = ArgIndices.begin();
848 *It != Operands; ++It, ++TheArg) {
849 assert(It != ArgIndices.end() && "GEP not handled??");
850 }
852 std::string NewName = I->getName();
853 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
854 NewName += "." + utostr(Operands[i]);
855 }
856 NewName += ".val";
857 TheArg->setName(NewName);
859 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
860 << "' of function '" << NF->getName() << "'\n");
862 // All of the uses must be load instructions. Replace them all with
863 // the argument specified by ArgNo.
864 while (!GEP->use_empty()) {
865 LoadInst *L = cast<LoadInst>(GEP->use_back());
866 L->replaceAllUsesWith(TheArg);
867 AA.replaceWithNewValue(L, TheArg);
868 L->eraseFromParent();
869 }
870 AA.deleteValue(GEP);
871 GEP->eraseFromParent();
872 }
873 }
875 // Increment I2 past all of the arguments added for this promoted pointer.
876 std::advance(I2, ArgIndices.size());
877 }
879 // Tell the alias analysis that the old function is about to disappear.
880 AA.replaceWithNewValue(F, NF);
883 NF_CGN->stealCalledFunctionsFrom(CG[F]);
885 // Now that the old function is dead, delete it. If there is a dangling
886 // reference to the CallgraphNode, just leave the dead function around for
887 // someone else to nuke.
888 CallGraphNode *CGN = CG[F];
889 if (CGN->getNumReferences() == 0)
890 delete CG.removeFunctionFromModule(CGN);
891 else
892 F->setLinkage(Function::ExternalLinkage);
894 return NF_CGN;
895 }