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 #include "llvm/Transforms/IPO.h"
33 #include "llvm/ADT/DepthFirstIterator.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/ADT/StringExtras.h"
36 #include "llvm/Analysis/AliasAnalysis.h"
37 #include "llvm/Analysis/CallGraph.h"
38 #include "llvm/Analysis/CallGraphSCCPass.h"
39 #include "llvm/IR/CFG.h"
40 #include "llvm/IR/CallSite.h"
41 #include "llvm/IR/Constants.h"
42 #include "llvm/IR/DataLayout.h"
43 #include "llvm/IR/DebugInfo.h"
44 #include "llvm/IR/DerivedTypes.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/LLVMContext.h"
47 #include "llvm/IR/Module.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include <set>
51 using namespace llvm;
52
53 #define DEBUG_TYPE "argpromotion"
54
55 STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted");
56 STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted");
57 STATISTIC(NumByValArgsPromoted , "Number of byval arguments promoted");
58 STATISTIC(NumArgumentsDead , "Number of dead pointer args eliminated");
59
60 namespace {
61 /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass.
62 ///
63 struct ArgPromotion : public CallGraphSCCPass {
getAnalysisUsage__anon77fa3f1a0111::ArgPromotion64 void getAnalysisUsage(AnalysisUsage &AU) const override {
65 AU.addRequired<AliasAnalysis>();
66 CallGraphSCCPass::getAnalysisUsage(AU);
67 }
68
69 bool runOnSCC(CallGraphSCC &SCC) override;
70 static char ID; // Pass identification, replacement for typeid
ArgPromotion__anon77fa3f1a0111::ArgPromotion71 explicit ArgPromotion(unsigned maxElements = 3)
72 : CallGraphSCCPass(ID), DL(nullptr), maxElements(maxElements) {
73 initializeArgPromotionPass(*PassRegistry::getPassRegistry());
74 }
75
76 /// A vector used to hold the indices of a single GEP instruction
77 typedef std::vector<uint64_t> IndicesVector;
78
79 const DataLayout *DL;
80 private:
81 bool isDenselyPacked(Type *type);
82 bool canPaddingBeAccessed(Argument *Arg);
83 CallGraphNode *PromoteArguments(CallGraphNode *CGN);
84 bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const;
85 CallGraphNode *DoPromotion(Function *F,
86 SmallPtrSetImpl<Argument*> &ArgsToPromote,
87 SmallPtrSetImpl<Argument*> &ByValArgsToTransform);
88
89 using llvm::Pass::doInitialization;
90 bool doInitialization(CallGraph &CG) override;
91 /// The maximum number of elements to expand, or 0 for unlimited.
92 unsigned maxElements;
93 DenseMap<const Function *, DISubprogram> FunctionDIs;
94 };
95 }
96
97 char ArgPromotion::ID = 0;
98 INITIALIZE_PASS_BEGIN(ArgPromotion, "argpromotion",
99 "Promote 'by reference' arguments to scalars", false, false)
INITIALIZE_AG_DEPENDENCY(AliasAnalysis)100 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
101 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
102 INITIALIZE_PASS_END(ArgPromotion, "argpromotion",
103 "Promote 'by reference' arguments to scalars", false, false)
104
105 Pass *llvm::createArgumentPromotionPass(unsigned maxElements) {
106 return new ArgPromotion(maxElements);
107 }
108
runOnSCC(CallGraphSCC & SCC)109 bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) {
110 bool Changed = false, LocalChange;
111
112 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
113 DL = DLP ? &DLP->getDataLayout() : nullptr;
114
115 do { // Iterate until we stop promoting from this SCC.
116 LocalChange = false;
117 // Attempt to promote arguments from all functions in this SCC.
118 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
119 if (CallGraphNode *CGN = PromoteArguments(*I)) {
120 LocalChange = true;
121 SCC.ReplaceNode(*I, CGN);
122 }
123 }
124 Changed |= LocalChange; // Remember that we changed something.
125 } while (LocalChange);
126
127 return Changed;
128 }
129
130 /// \brief Checks if a type could have padding bytes.
isDenselyPacked(Type * type)131 bool ArgPromotion::isDenselyPacked(Type *type) {
132
133 // There is no size information, so be conservative.
134 if (!type->isSized())
135 return false;
136
137 // If the alloc size is not equal to the storage size, then there are padding
138 // bytes. For x86_fp80 on x86-64, size: 80 alloc size: 128.
139 if (!DL || DL->getTypeSizeInBits(type) != DL->getTypeAllocSizeInBits(type))
140 return false;
141
142 if (!isa<CompositeType>(type))
143 return true;
144
145 // For homogenous sequential types, check for padding within members.
146 if (SequentialType *seqTy = dyn_cast<SequentialType>(type))
147 return isa<PointerType>(seqTy) || isDenselyPacked(seqTy->getElementType());
148
149 // Check for padding within and between elements of a struct.
150 StructType *StructTy = cast<StructType>(type);
151 const StructLayout *Layout = DL->getStructLayout(StructTy);
152 uint64_t StartPos = 0;
153 for (unsigned i = 0, E = StructTy->getNumElements(); i < E; ++i) {
154 Type *ElTy = StructTy->getElementType(i);
155 if (!isDenselyPacked(ElTy))
156 return false;
157 if (StartPos != Layout->getElementOffsetInBits(i))
158 return false;
159 StartPos += DL->getTypeAllocSizeInBits(ElTy);
160 }
161
162 return true;
163 }
164
165 /// \brief Checks if the padding bytes of an argument could be accessed.
canPaddingBeAccessed(Argument * arg)166 bool ArgPromotion::canPaddingBeAccessed(Argument *arg) {
167
168 assert(arg->hasByValAttr());
169
170 // Track all the pointers to the argument to make sure they are not captured.
171 SmallPtrSet<Value *, 16> PtrValues;
172 PtrValues.insert(arg);
173
174 // Track all of the stores.
175 SmallVector<StoreInst *, 16> Stores;
176
177 // Scan through the uses recursively to make sure the pointer is always used
178 // sanely.
179 SmallVector<Value *, 16> WorkList;
180 WorkList.insert(WorkList.end(), arg->user_begin(), arg->user_end());
181 while (!WorkList.empty()) {
182 Value *V = WorkList.back();
183 WorkList.pop_back();
184 if (isa<GetElementPtrInst>(V) || isa<PHINode>(V)) {
185 if (PtrValues.insert(V).second)
186 WorkList.insert(WorkList.end(), V->user_begin(), V->user_end());
187 } else if (StoreInst *Store = dyn_cast<StoreInst>(V)) {
188 Stores.push_back(Store);
189 } else if (!isa<LoadInst>(V)) {
190 return true;
191 }
192 }
193
194 // Check to make sure the pointers aren't captured
195 for (StoreInst *Store : Stores)
196 if (PtrValues.count(Store->getValueOperand()))
197 return true;
198
199 return false;
200 }
201
202 /// PromoteArguments - This method checks the specified function to see if there
203 /// are any promotable arguments and if it is safe to promote the function (for
204 /// example, all callers are direct). If safe to promote some arguments, it
205 /// calls the DoPromotion method.
206 ///
PromoteArguments(CallGraphNode * CGN)207 CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) {
208 Function *F = CGN->getFunction();
209
210 // Make sure that it is local to this module.
211 if (!F || !F->hasLocalLinkage()) return nullptr;
212
213 // First check: see if there are any pointer arguments! If not, quick exit.
214 SmallVector<Argument*, 16> PointerArgs;
215 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
216 if (I->getType()->isPointerTy())
217 PointerArgs.push_back(I);
218 if (PointerArgs.empty()) return nullptr;
219
220 // Second check: make sure that all callers are direct callers. We can't
221 // transform functions that have indirect callers. Also see if the function
222 // is self-recursive.
223 bool isSelfRecursive = false;
224 for (Use &U : F->uses()) {
225 CallSite CS(U.getUser());
226 // Must be a direct call.
227 if (CS.getInstruction() == nullptr || !CS.isCallee(&U)) return nullptr;
228
229 if (CS.getInstruction()->getParent()->getParent() == F)
230 isSelfRecursive = true;
231 }
232
233 // Don't promote arguments for variadic functions. Adding, removing, or
234 // changing non-pack parameters can change the classification of pack
235 // parameters. Frontends encode that classification at the call site in the
236 // IR, while in the callee the classification is determined dynamically based
237 // on the number of registers consumed so far.
238 if (F->isVarArg()) return nullptr;
239
240 // Check to see which arguments are promotable. If an argument is promotable,
241 // add it to ArgsToPromote.
242 SmallPtrSet<Argument*, 8> ArgsToPromote;
243 SmallPtrSet<Argument*, 8> ByValArgsToTransform;
244 for (unsigned i = 0, e = PointerArgs.size(); i != e; ++i) {
245 Argument *PtrArg = PointerArgs[i];
246 Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType();
247
248 // If this is a byval argument, and if the aggregate type is small, just
249 // pass the elements, which is always safe, if the passed value is densely
250 // packed or if we can prove the padding bytes are never accessed. This does
251 // not apply to inalloca.
252 bool isSafeToPromote =
253 PtrArg->hasByValAttr() &&
254 (isDenselyPacked(AgTy) || !canPaddingBeAccessed(PtrArg));
255 if (isSafeToPromote) {
256 if (StructType *STy = dyn_cast<StructType>(AgTy)) {
257 if (maxElements > 0 && STy->getNumElements() > maxElements) {
258 DEBUG(dbgs() << "argpromotion disable promoting argument '"
259 << PtrArg->getName() << "' because it would require adding more"
260 << " than " << maxElements << " arguments to the function.\n");
261 continue;
262 }
263
264 // If all the elements are single-value types, we can promote it.
265 bool AllSimple = true;
266 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
267 if (!STy->getElementType(i)->isSingleValueType()) {
268 AllSimple = false;
269 break;
270 }
271 }
272
273 // Safe to transform, don't even bother trying to "promote" it.
274 // Passing the elements as a scalar will allow scalarrepl to hack on
275 // the new alloca we introduce.
276 if (AllSimple) {
277 ByValArgsToTransform.insert(PtrArg);
278 continue;
279 }
280 }
281 }
282
283 // If the argument is a recursive type and we're in a recursive
284 // function, we could end up infinitely peeling the function argument.
285 if (isSelfRecursive) {
286 if (StructType *STy = dyn_cast<StructType>(AgTy)) {
287 bool RecursiveType = false;
288 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
289 if (STy->getElementType(i) == PtrArg->getType()) {
290 RecursiveType = true;
291 break;
292 }
293 }
294 if (RecursiveType)
295 continue;
296 }
297 }
298
299 // Otherwise, see if we can promote the pointer to its value.
300 if (isSafeToPromoteArgument(PtrArg, PtrArg->hasByValOrInAllocaAttr()))
301 ArgsToPromote.insert(PtrArg);
302 }
303
304 // No promotable pointer arguments.
305 if (ArgsToPromote.empty() && ByValArgsToTransform.empty())
306 return nullptr;
307
308 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform);
309 }
310
311 /// AllCallersPassInValidPointerForArgument - Return true if we can prove that
312 /// all callees pass in a valid pointer for the specified function argument.
AllCallersPassInValidPointerForArgument(Argument * Arg,const DataLayout * DL)313 static bool AllCallersPassInValidPointerForArgument(Argument *Arg,
314 const DataLayout *DL) {
315 Function *Callee = Arg->getParent();
316
317 unsigned ArgNo = Arg->getArgNo();
318
319 // Look at all call sites of the function. At this pointer we know we only
320 // have direct callees.
321 for (User *U : Callee->users()) {
322 CallSite CS(U);
323 assert(CS && "Should only have direct calls!");
324
325 if (!CS.getArgument(ArgNo)->isDereferenceablePointer(DL))
326 return false;
327 }
328 return true;
329 }
330
331 /// Returns true if Prefix is a prefix of longer. That means, Longer has a size
332 /// that is greater than or equal to the size of prefix, and each of the
333 /// elements in Prefix is the same as the corresponding elements in Longer.
334 ///
335 /// This means it also returns true when Prefix and Longer are equal!
IsPrefix(const ArgPromotion::IndicesVector & Prefix,const ArgPromotion::IndicesVector & Longer)336 static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix,
337 const ArgPromotion::IndicesVector &Longer) {
338 if (Prefix.size() > Longer.size())
339 return false;
340 return std::equal(Prefix.begin(), Prefix.end(), Longer.begin());
341 }
342
343
344 /// Checks if Indices, or a prefix of Indices, is in Set.
PrefixIn(const ArgPromotion::IndicesVector & Indices,std::set<ArgPromotion::IndicesVector> & Set)345 static bool PrefixIn(const ArgPromotion::IndicesVector &Indices,
346 std::set<ArgPromotion::IndicesVector> &Set) {
347 std::set<ArgPromotion::IndicesVector>::iterator Low;
348 Low = Set.upper_bound(Indices);
349 if (Low != Set.begin())
350 Low--;
351 // Low is now the last element smaller than or equal to Indices. This means
352 // it points to a prefix of Indices (possibly Indices itself), if such
353 // prefix exists.
354 //
355 // This load is safe if any prefix of its operands is safe to load.
356 return Low != Set.end() && IsPrefix(*Low, Indices);
357 }
358
359 /// Mark the given indices (ToMark) as safe in the given set of indices
360 /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there
361 /// is already a prefix of Indices in Safe, Indices are implicitely marked safe
362 /// already. Furthermore, any indices that Indices is itself a prefix of, are
363 /// removed from Safe (since they are implicitely safe because of Indices now).
MarkIndicesSafe(const ArgPromotion::IndicesVector & ToMark,std::set<ArgPromotion::IndicesVector> & Safe)364 static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark,
365 std::set<ArgPromotion::IndicesVector> &Safe) {
366 std::set<ArgPromotion::IndicesVector>::iterator Low;
367 Low = Safe.upper_bound(ToMark);
368 // Guard against the case where Safe is empty
369 if (Low != Safe.begin())
370 Low--;
371 // Low is now the last element smaller than or equal to Indices. This
372 // means it points to a prefix of Indices (possibly Indices itself), if
373 // such prefix exists.
374 if (Low != Safe.end()) {
375 if (IsPrefix(*Low, ToMark))
376 // If there is already a prefix of these indices (or exactly these
377 // indices) marked a safe, don't bother adding these indices
378 return;
379
380 // Increment Low, so we can use it as a "insert before" hint
381 ++Low;
382 }
383 // Insert
384 Low = Safe.insert(Low, ToMark);
385 ++Low;
386 // If there we're a prefix of longer index list(s), remove those
387 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end();
388 while (Low != End && IsPrefix(ToMark, *Low)) {
389 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low;
390 ++Low;
391 Safe.erase(Remove);
392 }
393 }
394
395 /// isSafeToPromoteArgument - As you might guess from the name of this method,
396 /// it checks to see if it is both safe and useful to promote the argument.
397 /// This method limits promotion of aggregates to only promote up to three
398 /// elements of the aggregate in order to avoid exploding the number of
399 /// arguments passed in.
isSafeToPromoteArgument(Argument * Arg,bool isByValOrInAlloca) const400 bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg,
401 bool isByValOrInAlloca) const {
402 typedef std::set<IndicesVector> GEPIndicesSet;
403
404 // Quick exit for unused arguments
405 if (Arg->use_empty())
406 return true;
407
408 // We can only promote this argument if all of the uses are loads, or are GEP
409 // instructions (with constant indices) that are subsequently loaded.
410 //
411 // Promoting the argument causes it to be loaded in the caller
412 // unconditionally. This is only safe if we can prove that either the load
413 // would have happened in the callee anyway (ie, there is a load in the entry
414 // block) or the pointer passed in at every call site is guaranteed to be
415 // valid.
416 // In the former case, invalid loads can happen, but would have happened
417 // anyway, in the latter case, invalid loads won't happen. This prevents us
418 // from introducing an invalid load that wouldn't have happened in the
419 // original code.
420 //
421 // This set will contain all sets of indices that are loaded in the entry
422 // block, and thus are safe to unconditionally load in the caller.
423 //
424 // This optimization is also safe for InAlloca parameters, because it verifies
425 // that the address isn't captured.
426 GEPIndicesSet SafeToUnconditionallyLoad;
427
428 // This set contains all the sets of indices that we are planning to promote.
429 // This makes it possible to limit the number of arguments added.
430 GEPIndicesSet ToPromote;
431
432 // If the pointer is always valid, any load with first index 0 is valid.
433 if (isByValOrInAlloca || AllCallersPassInValidPointerForArgument(Arg, DL))
434 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0));
435
436 // First, iterate the entry block and mark loads of (geps of) arguments as
437 // safe.
438 BasicBlock *EntryBlock = Arg->getParent()->begin();
439 // Declare this here so we can reuse it
440 IndicesVector Indices;
441 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end();
442 I != E; ++I)
443 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
444 Value *V = LI->getPointerOperand();
445 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
446 V = GEP->getPointerOperand();
447 if (V == Arg) {
448 // This load actually loads (part of) Arg? Check the indices then.
449 Indices.reserve(GEP->getNumIndices());
450 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
451 II != IE; ++II)
452 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II))
453 Indices.push_back(CI->getSExtValue());
454 else
455 // We found a non-constant GEP index for this argument? Bail out
456 // right away, can't promote this argument at all.
457 return false;
458
459 // Indices checked out, mark them as safe
460 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad);
461 Indices.clear();
462 }
463 } else if (V == Arg) {
464 // Direct loads are equivalent to a GEP with a single 0 index.
465 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad);
466 }
467 }
468
469 // Now, iterate all uses of the argument to see if there are any uses that are
470 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote.
471 SmallVector<LoadInst*, 16> Loads;
472 IndicesVector Operands;
473 for (Use &U : Arg->uses()) {
474 User *UR = U.getUser();
475 Operands.clear();
476 if (LoadInst *LI = dyn_cast<LoadInst>(UR)) {
477 // Don't hack volatile/atomic loads
478 if (!LI->isSimple()) return false;
479 Loads.push_back(LI);
480 // Direct loads are equivalent to a GEP with a zero index and then a load.
481 Operands.push_back(0);
482 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UR)) {
483 if (GEP->use_empty()) {
484 // Dead GEP's cause trouble later. Just remove them if we run into
485 // them.
486 getAnalysis<AliasAnalysis>().deleteValue(GEP);
487 GEP->eraseFromParent();
488 // TODO: This runs the above loop over and over again for dead GEPs
489 // Couldn't we just do increment the UI iterator earlier and erase the
490 // use?
491 return isSafeToPromoteArgument(Arg, isByValOrInAlloca);
492 }
493
494 // Ensure that all of the indices are constants.
495 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end();
496 i != e; ++i)
497 if (ConstantInt *C = dyn_cast<ConstantInt>(*i))
498 Operands.push_back(C->getSExtValue());
499 else
500 return false; // Not a constant operand GEP!
501
502 // Ensure that the only users of the GEP are load instructions.
503 for (User *GEPU : GEP->users())
504 if (LoadInst *LI = dyn_cast<LoadInst>(GEPU)) {
505 // Don't hack volatile/atomic loads
506 if (!LI->isSimple()) return false;
507 Loads.push_back(LI);
508 } else {
509 // Other uses than load?
510 return false;
511 }
512 } else {
513 return false; // Not a load or a GEP.
514 }
515
516 // Now, see if it is safe to promote this load / loads of this GEP. Loading
517 // is safe if Operands, or a prefix of Operands, is marked as safe.
518 if (!PrefixIn(Operands, SafeToUnconditionallyLoad))
519 return false;
520
521 // See if we are already promoting a load with these indices. If not, check
522 // to make sure that we aren't promoting too many elements. If so, nothing
523 // to do.
524 if (ToPromote.find(Operands) == ToPromote.end()) {
525 if (maxElements > 0 && ToPromote.size() == maxElements) {
526 DEBUG(dbgs() << "argpromotion not promoting argument '"
527 << Arg->getName() << "' because it would require adding more "
528 << "than " << maxElements << " arguments to the function.\n");
529 // We limit aggregate promotion to only promoting up to a fixed number
530 // of elements of the aggregate.
531 return false;
532 }
533 ToPromote.insert(std::move(Operands));
534 }
535 }
536
537 if (Loads.empty()) return true; // No users, this is a dead argument.
538
539 // Okay, now we know that the argument is only used by load instructions and
540 // it is safe to unconditionally perform all of them. Use alias analysis to
541 // check to see if the pointer is guaranteed to not be modified from entry of
542 // the function to each of the load instructions.
543
544 // Because there could be several/many load instructions, remember which
545 // blocks we know to be transparent to the load.
546 SmallPtrSet<BasicBlock*, 16> TranspBlocks;
547
548 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
549
550 for (unsigned i = 0, e = Loads.size(); i != e; ++i) {
551 // Check to see if the load is invalidated from the start of the block to
552 // the load itself.
553 LoadInst *Load = Loads[i];
554 BasicBlock *BB = Load->getParent();
555
556 AliasAnalysis::Location Loc = AA.getLocation(Load);
557 if (AA.canInstructionRangeModRef(BB->front(), *Load, Loc,
558 AliasAnalysis::Mod))
559 return false; // Pointer is invalidated!
560
561 // Now check every path from the entry block to the load for transparency.
562 // To do this, we perform a depth first search on the inverse CFG from the
563 // loading block.
564 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
565 BasicBlock *P = *PI;
566 for (BasicBlock *TranspBB : inverse_depth_first_ext(P, TranspBlocks))
567 if (AA.canBasicBlockModify(*TranspBB, Loc))
568 return false;
569 }
570 }
571
572 // If the path from the entry of the function to each load is free of
573 // instructions that potentially invalidate the load, we can make the
574 // transformation!
575 return true;
576 }
577
578 /// DoPromotion - This method actually performs the promotion of the specified
579 /// arguments, and returns the new function. At this point, we know that it's
580 /// safe to do so.
DoPromotion(Function * F,SmallPtrSetImpl<Argument * > & ArgsToPromote,SmallPtrSetImpl<Argument * > & ByValArgsToTransform)581 CallGraphNode *ArgPromotion::DoPromotion(Function *F,
582 SmallPtrSetImpl<Argument*> &ArgsToPromote,
583 SmallPtrSetImpl<Argument*> &ByValArgsToTransform) {
584
585 // Start by computing a new prototype for the function, which is the same as
586 // the old function, but has modified arguments.
587 FunctionType *FTy = F->getFunctionType();
588 std::vector<Type*> Params;
589
590 typedef std::set<IndicesVector> ScalarizeTable;
591
592 // ScalarizedElements - If we are promoting a pointer that has elements
593 // accessed out of it, keep track of which elements are accessed so that we
594 // can add one argument for each.
595 //
596 // Arguments that are directly loaded will have a zero element value here, to
597 // handle cases where there are both a direct load and GEP accesses.
598 //
599 std::map<Argument*, ScalarizeTable> ScalarizedElements;
600
601 // OriginalLoads - Keep track of a representative load instruction from the
602 // original function so that we can tell the alias analysis implementation
603 // what the new GEP/Load instructions we are inserting look like.
604 // We need to keep the original loads for each argument and the elements
605 // of the argument that are accessed.
606 std::map<std::pair<Argument*, IndicesVector>, LoadInst*> OriginalLoads;
607
608 // Attribute - Keep track of the parameter attributes for the arguments
609 // that we are *not* promoting. For the ones that we do promote, the parameter
610 // attributes are lost
611 SmallVector<AttributeSet, 8> AttributesVec;
612 const AttributeSet &PAL = F->getAttributes();
613
614 // Add any return attributes.
615 if (PAL.hasAttributes(AttributeSet::ReturnIndex))
616 AttributesVec.push_back(AttributeSet::get(F->getContext(),
617 PAL.getRetAttributes()));
618
619 // First, determine the new argument list
620 unsigned ArgIndex = 1;
621 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
622 ++I, ++ArgIndex) {
623 if (ByValArgsToTransform.count(I)) {
624 // Simple byval argument? Just add all the struct element types.
625 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
626 StructType *STy = cast<StructType>(AgTy);
627 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
628 Params.push_back(STy->getElementType(i));
629 ++NumByValArgsPromoted;
630 } else if (!ArgsToPromote.count(I)) {
631 // Unchanged argument
632 Params.push_back(I->getType());
633 AttributeSet attrs = PAL.getParamAttributes(ArgIndex);
634 if (attrs.hasAttributes(ArgIndex)) {
635 AttrBuilder B(attrs, ArgIndex);
636 AttributesVec.
637 push_back(AttributeSet::get(F->getContext(), Params.size(), B));
638 }
639 } else if (I->use_empty()) {
640 // Dead argument (which are always marked as promotable)
641 ++NumArgumentsDead;
642 } else {
643 // Okay, this is being promoted. This means that the only uses are loads
644 // or GEPs which are only used by loads
645
646 // In this table, we will track which indices are loaded from the argument
647 // (where direct loads are tracked as no indices).
648 ScalarizeTable &ArgIndices = ScalarizedElements[I];
649 for (User *U : I->users()) {
650 Instruction *UI = cast<Instruction>(U);
651 assert(isa<LoadInst>(UI) || isa<GetElementPtrInst>(UI));
652 IndicesVector Indices;
653 Indices.reserve(UI->getNumOperands() - 1);
654 // Since loads will only have a single operand, and GEPs only a single
655 // non-index operand, this will record direct loads without any indices,
656 // and gep+loads with the GEP indices.
657 for (User::op_iterator II = UI->op_begin() + 1, IE = UI->op_end();
658 II != IE; ++II)
659 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
660 // GEPs with a single 0 index can be merged with direct loads
661 if (Indices.size() == 1 && Indices.front() == 0)
662 Indices.clear();
663 ArgIndices.insert(Indices);
664 LoadInst *OrigLoad;
665 if (LoadInst *L = dyn_cast<LoadInst>(UI))
666 OrigLoad = L;
667 else
668 // Take any load, we will use it only to update Alias Analysis
669 OrigLoad = cast<LoadInst>(UI->user_back());
670 OriginalLoads[std::make_pair(I, Indices)] = OrigLoad;
671 }
672
673 // Add a parameter to the function for each element passed in.
674 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
675 E = ArgIndices.end(); SI != E; ++SI) {
676 // not allowed to dereference ->begin() if size() is 0
677 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), *SI));
678 assert(Params.back());
679 }
680
681 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty())
682 ++NumArgumentsPromoted;
683 else
684 ++NumAggregatesPromoted;
685 }
686 }
687
688 // Add any function attributes.
689 if (PAL.hasAttributes(AttributeSet::FunctionIndex))
690 AttributesVec.push_back(AttributeSet::get(FTy->getContext(),
691 PAL.getFnAttributes()));
692
693 Type *RetTy = FTy->getReturnType();
694
695 // Construct the new function type using the new arguments.
696 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
697
698 // Create the new function body and insert it into the module.
699 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
700 NF->copyAttributesFrom(F);
701
702 // Patch the pointer to LLVM function in debug info descriptor.
703 auto DI = FunctionDIs.find(F);
704 if (DI != FunctionDIs.end()) {
705 DISubprogram SP = DI->second;
706 SP.replaceFunction(NF);
707 // Ensure the map is updated so it can be reused on subsequent argument
708 // promotions of the same function.
709 FunctionDIs.erase(DI);
710 FunctionDIs[NF] = SP;
711 }
712
713 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n"
714 << "From: " << *F);
715
716 // Recompute the parameter attributes list based on the new arguments for
717 // the function.
718 NF->setAttributes(AttributeSet::get(F->getContext(), AttributesVec));
719 AttributesVec.clear();
720
721 F->getParent()->getFunctionList().insert(F, NF);
722 NF->takeName(F);
723
724 // Get the alias analysis information that we need to update to reflect our
725 // changes.
726 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
727
728 // Get the callgraph information that we need to update to reflect our
729 // changes.
730 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
731
732 // Get a new callgraph node for NF.
733 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
734
735 // Loop over all of the callers of the function, transforming the call sites
736 // to pass in the loaded pointers.
737 //
738 SmallVector<Value*, 16> Args;
739 while (!F->use_empty()) {
740 CallSite CS(F->user_back());
741 assert(CS.getCalledFunction() == F);
742 Instruction *Call = CS.getInstruction();
743 const AttributeSet &CallPAL = CS.getAttributes();
744
745 // Add any return attributes.
746 if (CallPAL.hasAttributes(AttributeSet::ReturnIndex))
747 AttributesVec.push_back(AttributeSet::get(F->getContext(),
748 CallPAL.getRetAttributes()));
749
750 // Loop over the operands, inserting GEP and loads in the caller as
751 // appropriate.
752 CallSite::arg_iterator AI = CS.arg_begin();
753 ArgIndex = 1;
754 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
755 I != E; ++I, ++AI, ++ArgIndex)
756 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
757 Args.push_back(*AI); // Unmodified argument
758
759 if (CallPAL.hasAttributes(ArgIndex)) {
760 AttrBuilder B(CallPAL, ArgIndex);
761 AttributesVec.
762 push_back(AttributeSet::get(F->getContext(), Args.size(), B));
763 }
764 } else if (ByValArgsToTransform.count(I)) {
765 // Emit a GEP and load for each element of the struct.
766 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
767 StructType *STy = cast<StructType>(AgTy);
768 Value *Idxs[2] = {
769 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
770 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
771 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
772 Value *Idx = GetElementPtrInst::Create(*AI, Idxs,
773 (*AI)->getName()+"."+utostr(i),
774 Call);
775 // TODO: Tell AA about the new values?
776 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
777 }
778 } else if (!I->use_empty()) {
779 // Non-dead argument: insert GEPs and loads as appropriate.
780 ScalarizeTable &ArgIndices = ScalarizedElements[I];
781 // Store the Value* version of the indices in here, but declare it now
782 // for reuse.
783 std::vector<Value*> Ops;
784 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
785 E = ArgIndices.end(); SI != E; ++SI) {
786 Value *V = *AI;
787 LoadInst *OrigLoad = OriginalLoads[std::make_pair(I, *SI)];
788 if (!SI->empty()) {
789 Ops.reserve(SI->size());
790 Type *ElTy = V->getType();
791 for (IndicesVector::const_iterator II = SI->begin(),
792 IE = SI->end(); II != IE; ++II) {
793 // Use i32 to index structs, and i64 for others (pointers/arrays).
794 // This satisfies GEP constraints.
795 Type *IdxTy = (ElTy->isStructTy() ?
796 Type::getInt32Ty(F->getContext()) :
797 Type::getInt64Ty(F->getContext()));
798 Ops.push_back(ConstantInt::get(IdxTy, *II));
799 // Keep track of the type we're currently indexing.
800 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
801 }
802 // And create a GEP to extract those indices.
803 V = GetElementPtrInst::Create(V, Ops, V->getName()+".idx", Call);
804 Ops.clear();
805 AA.copyValue(OrigLoad->getOperand(0), V);
806 }
807 // Since we're replacing a load make sure we take the alignment
808 // of the previous load.
809 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
810 newLoad->setAlignment(OrigLoad->getAlignment());
811 // Transfer the AA info too.
812 AAMDNodes AAInfo;
813 OrigLoad->getAAMetadata(AAInfo);
814 newLoad->setAAMetadata(AAInfo);
815
816 Args.push_back(newLoad);
817 AA.copyValue(OrigLoad, Args.back());
818 }
819 }
820
821 // Push any varargs arguments on the list.
822 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
823 Args.push_back(*AI);
824 if (CallPAL.hasAttributes(ArgIndex)) {
825 AttrBuilder B(CallPAL, ArgIndex);
826 AttributesVec.
827 push_back(AttributeSet::get(F->getContext(), Args.size(), B));
828 }
829 }
830
831 // Add any function attributes.
832 if (CallPAL.hasAttributes(AttributeSet::FunctionIndex))
833 AttributesVec.push_back(AttributeSet::get(Call->getContext(),
834 CallPAL.getFnAttributes()));
835
836 Instruction *New;
837 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
838 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
839 Args, "", Call);
840 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
841 cast<InvokeInst>(New)->setAttributes(AttributeSet::get(II->getContext(),
842 AttributesVec));
843 } else {
844 New = CallInst::Create(NF, Args, "", Call);
845 cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
846 cast<CallInst>(New)->setAttributes(AttributeSet::get(New->getContext(),
847 AttributesVec));
848 if (cast<CallInst>(Call)->isTailCall())
849 cast<CallInst>(New)->setTailCall();
850 }
851 New->setDebugLoc(Call->getDebugLoc());
852 Args.clear();
853 AttributesVec.clear();
854
855 // Update the alias analysis implementation to know that we are replacing
856 // the old call with a new one.
857 AA.replaceWithNewValue(Call, New);
858
859 // Update the callgraph to know that the callsite has been transformed.
860 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
861 CalleeNode->replaceCallEdge(Call, New, NF_CGN);
862
863 if (!Call->use_empty()) {
864 Call->replaceAllUsesWith(New);
865 New->takeName(Call);
866 }
867
868 // Finally, remove the old call from the program, reducing the use-count of
869 // F.
870 Call->eraseFromParent();
871 }
872
873 // Since we have now created the new function, splice the body of the old
874 // function right into the new function, leaving the old rotting hulk of the
875 // function empty.
876 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
877
878 // Loop over the argument list, transferring uses of the old arguments over to
879 // the new arguments, also transferring over the names as well.
880 //
881 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
882 I2 = NF->arg_begin(); I != E; ++I) {
883 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
884 // If this is an unmodified argument, move the name and users over to the
885 // new version.
886 I->replaceAllUsesWith(I2);
887 I2->takeName(I);
888 AA.replaceWithNewValue(I, I2);
889 ++I2;
890 continue;
891 }
892
893 if (ByValArgsToTransform.count(I)) {
894 // In the callee, we create an alloca, and store each of the new incoming
895 // arguments into the alloca.
896 Instruction *InsertPt = NF->begin()->begin();
897
898 // Just add all the struct element types.
899 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
900 Value *TheAlloca = new AllocaInst(AgTy, nullptr, "", InsertPt);
901 StructType *STy = cast<StructType>(AgTy);
902 Value *Idxs[2] = {
903 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
904
905 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
906 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
907 Value *Idx =
908 GetElementPtrInst::Create(TheAlloca, Idxs,
909 TheAlloca->getName()+"."+Twine(i),
910 InsertPt);
911 I2->setName(I->getName()+"."+Twine(i));
912 new StoreInst(I2++, Idx, InsertPt);
913 }
914
915 // Anything that used the arg should now use the alloca.
916 I->replaceAllUsesWith(TheAlloca);
917 TheAlloca->takeName(I);
918 AA.replaceWithNewValue(I, TheAlloca);
919
920 // If the alloca is used in a call, we must clear the tail flag since
921 // the callee now uses an alloca from the caller.
922 for (User *U : TheAlloca->users()) {
923 CallInst *Call = dyn_cast<CallInst>(U);
924 if (!Call)
925 continue;
926 Call->setTailCall(false);
927 }
928 continue;
929 }
930
931 if (I->use_empty()) {
932 AA.deleteValue(I);
933 continue;
934 }
935
936 // Otherwise, if we promoted this argument, then all users are load
937 // instructions (or GEPs with only load users), and all loads should be
938 // using the new argument that we added.
939 ScalarizeTable &ArgIndices = ScalarizedElements[I];
940
941 while (!I->use_empty()) {
942 if (LoadInst *LI = dyn_cast<LoadInst>(I->user_back())) {
943 assert(ArgIndices.begin()->empty() &&
944 "Load element should sort to front!");
945 I2->setName(I->getName()+".val");
946 LI->replaceAllUsesWith(I2);
947 AA.replaceWithNewValue(LI, I2);
948 LI->eraseFromParent();
949 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
950 << "' in function '" << F->getName() << "'\n");
951 } else {
952 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->user_back());
953 IndicesVector Operands;
954 Operands.reserve(GEP->getNumIndices());
955 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
956 II != IE; ++II)
957 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
958
959 // GEPs with a single 0 index can be merged with direct loads
960 if (Operands.size() == 1 && Operands.front() == 0)
961 Operands.clear();
962
963 Function::arg_iterator TheArg = I2;
964 for (ScalarizeTable::iterator It = ArgIndices.begin();
965 *It != Operands; ++It, ++TheArg) {
966 assert(It != ArgIndices.end() && "GEP not handled??");
967 }
968
969 std::string NewName = I->getName();
970 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
971 NewName += "." + utostr(Operands[i]);
972 }
973 NewName += ".val";
974 TheArg->setName(NewName);
975
976 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
977 << "' of function '" << NF->getName() << "'\n");
978
979 // All of the uses must be load instructions. Replace them all with
980 // the argument specified by ArgNo.
981 while (!GEP->use_empty()) {
982 LoadInst *L = cast<LoadInst>(GEP->user_back());
983 L->replaceAllUsesWith(TheArg);
984 AA.replaceWithNewValue(L, TheArg);
985 L->eraseFromParent();
986 }
987 AA.deleteValue(GEP);
988 GEP->eraseFromParent();
989 }
990 }
991
992 // Increment I2 past all of the arguments added for this promoted pointer.
993 std::advance(I2, ArgIndices.size());
994 }
995
996 // Tell the alias analysis that the old function is about to disappear.
997 AA.replaceWithNewValue(F, NF);
998
999
1000 NF_CGN->stealCalledFunctionsFrom(CG[F]);
1001
1002 // Now that the old function is dead, delete it. If there is a dangling
1003 // reference to the CallgraphNode, just leave the dead function around for
1004 // someone else to nuke.
1005 CallGraphNode *CGN = CG[F];
1006 if (CGN->getNumReferences() == 0)
1007 delete CG.removeFunctionFromModule(CGN);
1008 else
1009 F->setLinkage(Function::ExternalLinkage);
1010
1011 return NF_CGN;
1012 }
1013
doInitialization(CallGraph & CG)1014 bool ArgPromotion::doInitialization(CallGraph &CG) {
1015 FunctionDIs = makeSubprogramMap(CG.getModule());
1016 return CallGraphSCCPass::doInitialization(CG);
1017 }
1018