1 //===- MergeFunctions.cpp - Merge identical functions ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass looks for equivalent functions that are mergable and folds them.
12 // A hash is computed from the function, based on its type and number of
15 // Once all hashes are computed, we perform an expensive equality comparison
16 // on each function pair. This takes n^2/2 comparisons per bucket, so it's
17 // important that the hash function be high quality. The equality comparison
18 // iterates through each instruction in each basic block.
20 // When a match is found the functions are folded. If both functions are
21 // overridable, we move the functionality into a new internal function and
22 // leave two overridable thunks to it.
24 //===----------------------------------------------------------------------===//
28 // * virtual functions.
30 // Many functions have their address taken by the virtual function table for
31 // the object they belong to. However, as long as it's only used for a lookup
32 // and call, this is irrelevant, and we'd like to fold such functions.
34 // * switch from n^2 pair-wise comparisons to an n-way comparison for each
37 // * be smarter about bitcasts.
39 // In order to fold functions, we will sometimes add either bitcast instructions
40 // or bitcast constant expressions. Unfortunately, this can confound further
41 // analysis since the two functions differ where one has a bitcast and the
42 // other doesn't. We should learn to look through bitcasts.
44 //===----------------------------------------------------------------------===//
46 #include "llvm/Transforms/IPO.h"
47 #include "llvm/ADT/DenseSet.h"
48 #include "llvm/ADT/FoldingSet.h"
49 #include "llvm/ADT/STLExtras.h"
50 #include "llvm/ADT/SmallSet.h"
51 #include "llvm/ADT/Statistic.h"
52 #include "llvm/IR/CallSite.h"
53 #include "llvm/IR/Constants.h"
54 #include "llvm/IR/DataLayout.h"
55 #include "llvm/IR/IRBuilder.h"
56 #include "llvm/IR/InlineAsm.h"
57 #include "llvm/IR/Instructions.h"
58 #include "llvm/IR/LLVMContext.h"
59 #include "llvm/IR/Module.h"
60 #include "llvm/IR/Operator.h"
61 #include "llvm/IR/ValueHandle.h"
62 #include "llvm/Pass.h"
63 #include "llvm/Support/Debug.h"
64 #include "llvm/Support/ErrorHandling.h"
65 #include "llvm/Support/raw_ostream.h"
69 #define DEBUG_TYPE "mergefunc"
71 STATISTIC(NumFunctionsMerged, "Number of functions merged");
72 STATISTIC(NumThunksWritten, "Number of thunks generated");
73 STATISTIC(NumAliasesWritten, "Number of aliases generated");
74 STATISTIC(NumDoubleWeak, "Number of new functions created");
76 /// Returns the type id for a type to be hashed. We turn pointer types into
77 /// integers here because the actual compare logic below considers pointers and
78 /// integers of the same size as equal.
79 static Type::TypeID getTypeIDForHash(Type *Ty) {
80 if (Ty->isPointerTy())
81 return Type::IntegerTyID;
82 return Ty->getTypeID();
85 /// Creates a hash-code for the function which is the same for any two
86 /// functions that will compare equal, without looking at the instructions
87 /// inside the function.
88 static unsigned profileFunction(const Function *F) {
89 FunctionType *FTy = F->getFunctionType();
92 ID.AddInteger(F->size());
93 ID.AddInteger(F->getCallingConv());
94 ID.AddBoolean(F->hasGC());
95 ID.AddBoolean(FTy->isVarArg());
96 ID.AddInteger(getTypeIDForHash(FTy->getReturnType()));
97 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
98 ID.AddInteger(getTypeIDForHash(FTy->getParamType(i)));
99 return ID.ComputeHash();
104 /// ComparableFunction - A struct that pairs together functions with a
105 /// DataLayout so that we can keep them together as elements in the DenseSet.
106 class ComparableFunction {
108 static const ComparableFunction EmptyKey;
109 static const ComparableFunction TombstoneKey;
110 static DataLayout * const LookupOnly;
112 ComparableFunction(Function *Func, const DataLayout *DL)
113 : Func(Func), Hash(profileFunction(Func)), DL(DL) {}
115 Function *getFunc() const { return Func; }
116 unsigned getHash() const { return Hash; }
117 const DataLayout *getDataLayout() const { return DL; }
119 // Drops AssertingVH reference to the function. Outside of debug mode, this
123 "Attempted to release function twice, or release empty/tombstone!");
128 explicit ComparableFunction(unsigned Hash)
129 : Func(nullptr), Hash(Hash), DL(nullptr) {}
131 AssertingVH<Function> Func;
133 const DataLayout *DL;
136 const ComparableFunction ComparableFunction::EmptyKey = ComparableFunction(0);
137 const ComparableFunction ComparableFunction::TombstoneKey =
138 ComparableFunction(1);
139 DataLayout *const ComparableFunction::LookupOnly = (DataLayout*)(-1);
145 struct DenseMapInfo<ComparableFunction> {
146 static ComparableFunction getEmptyKey() {
147 return ComparableFunction::EmptyKey;
149 static ComparableFunction getTombstoneKey() {
150 return ComparableFunction::TombstoneKey;
152 static unsigned getHashValue(const ComparableFunction &CF) {
155 static bool isEqual(const ComparableFunction &LHS,
156 const ComparableFunction &RHS);
162 /// FunctionComparator - Compares two functions to determine whether or not
163 /// they will generate machine code with the same behaviour. DataLayout is
164 /// used if available. The comparator always fails conservatively (erring on the
165 /// side of claiming that two functions are different).
166 class FunctionComparator {
168 FunctionComparator(const DataLayout *DL, const Function *F1,
170 : F1(F1), F2(F2), DL(DL) {}
172 /// Test whether the two functions have equivalent behaviour.
176 /// Test whether two basic blocks have equivalent behaviour.
177 bool compare(const BasicBlock *BB1, const BasicBlock *BB2);
179 /// Constants comparison.
180 /// Its analog to lexicographical comparison between hypothetical numbers
182 /// <bitcastability-trait><raw-bit-contents>
184 /// 1. Bitcastability.
185 /// Check whether L's type could be losslessly bitcasted to R's type.
186 /// On this stage method, in case when lossless bitcast is not possible
187 /// method returns -1 or 1, thus also defining which type is greater in
188 /// context of bitcastability.
189 /// Stage 0: If types are equal in terms of cmpTypes, then we can go straight
190 /// to the contents comparison.
191 /// If types differ, remember types comparison result and check
192 /// whether we still can bitcast types.
193 /// Stage 1: Types that satisfies isFirstClassType conditions are always
194 /// greater then others.
195 /// Stage 2: Vector is greater then non-vector.
196 /// If both types are vectors, then vector with greater bitwidth is
198 /// If both types are vectors with the same bitwidth, then types
199 /// are bitcastable, and we can skip other stages, and go to contents
201 /// Stage 3: Pointer types are greater than non-pointers. If both types are
202 /// pointers of the same address space - go to contents comparison.
203 /// Different address spaces: pointer with greater address space is
205 /// Stage 4: Types are neither vectors, nor pointers. And they differ.
206 /// We don't know how to bitcast them. So, we better don't do it,
207 /// and return types comparison result (so it determines the
208 /// relationship among constants we don't know how to bitcast).
210 /// Just for clearance, let's see how the set of constants could look
211 /// on single dimension axis:
213 /// [NFCT], [FCT, "others"], [FCT, pointers], [FCT, vectors]
214 /// Where: NFCT - Not a FirstClassType
215 /// FCT - FirstClassTyp:
217 /// 2. Compare raw contents.
218 /// It ignores types on this stage and only compares bits from L and R.
219 /// Returns 0, if L and R has equivalent contents.
220 /// -1 or 1 if values are different.
222 /// 2.1. If contents are numbers, compare numbers.
223 /// Ints with greater bitwidth are greater. Ints with same bitwidths
224 /// compared by their contents.
225 /// 2.2. "And so on". Just to avoid discrepancies with comments
226 /// perhaps it would be better to read the implementation itself.
227 /// 3. And again about overall picture. Let's look back at how the ordered set
228 /// of constants will look like:
229 /// [NFCT], [FCT, "others"], [FCT, pointers], [FCT, vectors]
231 /// Now look, what could be inside [FCT, "others"], for example:
232 /// [FCT, "others"] =
234 /// [double 0.1], [double 1.23],
235 /// [i32 1], [i32 2],
236 /// { double 1.0 }, ; StructTyID, NumElements = 1
237 /// { i32 1 }, ; StructTyID, NumElements = 1
238 /// { double 1, i32 1 }, ; StructTyID, NumElements = 2
239 /// { i32 1, double 1 } ; StructTyID, NumElements = 2
242 /// Let's explain the order. Float numbers will be less than integers, just
243 /// because of cmpType terms: FloatTyID < IntegerTyID.
244 /// Floats (with same fltSemantics) are sorted according to their value.
245 /// Then you can see integers, and they are, like a floats,
246 /// could be easy sorted among each others.
247 /// The structures. Structures are grouped at the tail, again because of their
248 /// TypeID: StructTyID > IntegerTyID > FloatTyID.
249 /// Structures with greater number of elements are greater. Structures with
250 /// greater elements going first are greater.
251 /// The same logic with vectors, arrays and other possible complex types.
253 /// Bitcastable constants.
254 /// Let's assume, that some constant, belongs to some group of
255 /// "so-called-equal" values with different types, and at the same time
256 /// belongs to another group of constants with equal types
257 /// and "really" equal values.
259 /// Now, prove that this is impossible:
261 /// If constant A with type TyA is bitcastable to B with type TyB, then:
262 /// 1. All constants with equal types to TyA, are bitcastable to B. Since
263 /// those should be vectors (if TyA is vector), pointers
264 /// (if TyA is pointer), or else (if TyA equal to TyB), those types should
266 /// 2. All constants with non-equal, but bitcastable types to TyA, are
267 /// bitcastable to B.
268 /// Once again, just because we allow it to vectors and pointers only.
269 /// This statement could be expanded as below:
270 /// 2.1. All vectors with equal bitwidth to vector A, has equal bitwidth to
271 /// vector B, and thus bitcastable to B as well.
272 /// 2.2. All pointers of the same address space, no matter what they point to,
273 /// bitcastable. So if C is pointer, it could be bitcasted to A and to B.
274 /// So any constant equal or bitcastable to A is equal or bitcastable to B.
277 /// In another words, for pointers and vectors, we ignore top-level type and
278 /// look at their particular properties (bit-width for vectors, and
279 /// address space for pointers).
280 /// If these properties are equal - compare their contents.
281 int cmpConstants(const Constant *L, const Constant *R);
283 /// Assign or look up previously assigned numbers for the two values, and
284 /// return whether the numbers are equal. Numbers are assigned in the order
286 /// Comparison order:
287 /// Stage 0: Value that is function itself is always greater then others.
288 /// If left and right values are references to their functions, then
290 /// Stage 1: Constants are greater than non-constants.
291 /// If both left and right are constants, then the result of
292 /// cmpConstants is used as cmpValues result.
293 /// Stage 2: InlineAsm instances are greater than others. If both left and
294 /// right are InlineAsm instances, InlineAsm* pointers casted to
295 /// integers and compared as numbers.
296 /// Stage 3: For all other cases we compare order we meet these values in
297 /// their functions. If right value was met first during scanning,
298 /// then left value is greater.
299 /// In another words, we compare serial numbers, for more details
300 /// see comments for sn_mapL and sn_mapR.
301 int cmpValues(const Value *L, const Value *R);
303 bool enumerate(const Value *V1, const Value *V2) {
304 return cmpValues(V1, V2) == 0;
307 /// Compare two Instructions for equivalence, similar to
308 /// Instruction::isSameOperationAs but with modifications to the type
310 /// Stages are listed in "most significant stage first" order:
311 /// On each stage below, we do comparison between some left and right
312 /// operation parts. If parts are non-equal, we assign parts comparison
313 /// result to the operation comparison result and exit from method.
314 /// Otherwise we proceed to the next stage.
316 /// 1. Operations opcodes. Compared as numbers.
317 /// 2. Number of operands.
318 /// 3. Operation types. Compared with cmpType method.
319 /// 4. Compare operation subclass optional data as stream of bytes:
320 /// just convert it to integers and call cmpNumbers.
321 /// 5. Compare in operation operand types with cmpType in
322 /// most significant operand first order.
323 /// 6. Last stage. Check operations for some specific attributes.
324 /// For example, for Load it would be:
325 /// 6.1.Load: volatile (as boolean flag)
326 /// 6.2.Load: alignment (as integer numbers)
327 /// 6.3.Load: synch-scope (as integer numbers)
328 /// 6.4.Load: range metadata (as integer numbers)
329 /// On this stage its better to see the code, since its not more than 10-15
330 /// strings for particular instruction, and could change sometimes.
331 int cmpOperation(const Instruction *L, const Instruction *R) const;
333 bool isEquivalentOperation(const Instruction *I1,
334 const Instruction *I2) const {
335 return cmpOperation(I1, I2) == 0;
338 /// Compare two GEPs for equivalent pointer arithmetic.
339 /// Parts to be compared for each comparison stage,
340 /// most significant stage first:
341 /// 1. Address space. As numbers.
342 /// 2. Constant offset, (if "DataLayout *DL" field is not NULL,
343 /// using GEPOperator::accumulateConstantOffset method).
344 /// 3. Pointer operand type (using cmpType method).
345 /// 4. Number of operands.
346 /// 5. Compare operands, using cmpValues method.
347 int cmpGEP(const GEPOperator *GEPL, const GEPOperator *GEPR);
348 int cmpGEP(const GetElementPtrInst *GEPL, const GetElementPtrInst *GEPR) {
349 return cmpGEP(cast<GEPOperator>(GEPL), cast<GEPOperator>(GEPR));
352 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2) {
353 return cmpGEP(GEP1, GEP2) == 0;
355 bool isEquivalentGEP(const GetElementPtrInst *GEP1,
356 const GetElementPtrInst *GEP2) {
357 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2));
360 /// cmpType - compares two types,
361 /// defines total ordering among the types set.
364 /// 0 if types are equal,
365 /// -1 if Left is less than Right,
366 /// +1 if Left is greater than Right.
369 /// Comparison is broken onto stages. Like in lexicographical comparison
370 /// stage coming first has higher priority.
371 /// On each explanation stage keep in mind total ordering properties.
373 /// 0. Before comparison we coerce pointer types of 0 address space to
375 /// We also don't bother with same type at left and right, so
376 /// just return 0 in this case.
378 /// 1. If types are of different kind (different type IDs).
379 /// Return result of type IDs comparison, treating them as numbers.
380 /// 2. If types are vectors or integers, compare Type* values as numbers.
381 /// 3. Types has same ID, so check whether they belongs to the next group:
390 /// If so - return 0, yes - we can treat these types as equal only because
391 /// their IDs are same.
392 /// 4. If Left and Right are pointers, return result of address space
393 /// comparison (numbers comparison). We can treat pointer types of same
394 /// address space as equal.
395 /// 5. If types are complex.
396 /// Then both Left and Right are to be expanded and their element types will
397 /// be checked with the same way. If we get Res != 0 on some stage, return it.
398 /// Otherwise return 0.
399 /// 6. For all other cases put llvm_unreachable.
400 int cmpType(Type *TyL, Type *TyR) const;
402 bool isEquivalentType(Type *Ty1, Type *Ty2) const {
403 return cmpType(Ty1, Ty2) == 0;
406 int cmpNumbers(uint64_t L, uint64_t R) const;
408 int cmpAPInt(const APInt &L, const APInt &R) const;
409 int cmpAPFloat(const APFloat &L, const APFloat &R) const;
410 int cmpStrings(StringRef L, StringRef R) const;
411 int cmpAttrs(const AttributeSet L, const AttributeSet R) const;
413 // The two functions undergoing comparison.
414 const Function *F1, *F2;
416 const DataLayout *DL;
418 /// Assign serial numbers to values from left function, and values from
421 /// Being comparing functions we need to compare values we meet at left and
423 /// Its easy to sort things out for external values. It just should be
424 /// the same value at left and right.
425 /// But for local values (those were introduced inside function body)
426 /// we have to ensure they were introduced at exactly the same place,
427 /// and plays the same role.
428 /// Let's assign serial number to each value when we meet it first time.
429 /// Values that were met at same place will be with same serial numbers.
430 /// In this case it would be good to explain few points about values assigned
431 /// to BBs and other ways of implementation (see below).
433 /// 1. Safety of BB reordering.
434 /// It's safe to change the order of BasicBlocks in function.
435 /// Relationship with other functions and serial numbering will not be
436 /// changed in this case.
437 /// As follows from FunctionComparator::compare(), we do CFG walk: we start
438 /// from the entry, and then take each terminator. So it doesn't matter how in
439 /// fact BBs are ordered in function. And since cmpValues are called during
440 /// this walk, the numbering depends only on how BBs located inside the CFG.
441 /// So the answer is - yes. We will get the same numbering.
443 /// 2. Impossibility to use dominance properties of values.
444 /// If we compare two instruction operands: first is usage of local
445 /// variable AL from function FL, and second is usage of local variable AR
446 /// from FR, we could compare their origins and check whether they are
447 /// defined at the same place.
448 /// But, we are still not able to compare operands of PHI nodes, since those
449 /// could be operands from further BBs we didn't scan yet.
450 /// So it's impossible to use dominance properties in general.
451 DenseMap<const Value*, int> sn_mapL, sn_mapR;
456 int FunctionComparator::cmpNumbers(uint64_t L, uint64_t R) const {
457 if (L < R) return -1;
462 int FunctionComparator::cmpAPInt(const APInt &L, const APInt &R) const {
463 if (int Res = cmpNumbers(L.getBitWidth(), R.getBitWidth()))
465 if (L.ugt(R)) return 1;
466 if (R.ugt(L)) return -1;
470 int FunctionComparator::cmpAPFloat(const APFloat &L, const APFloat &R) const {
471 if (int Res = cmpNumbers((uint64_t)&L.getSemantics(),
472 (uint64_t)&R.getSemantics()))
474 return cmpAPInt(L.bitcastToAPInt(), R.bitcastToAPInt());
477 int FunctionComparator::cmpStrings(StringRef L, StringRef R) const {
478 // Prevent heavy comparison, compare sizes first.
479 if (int Res = cmpNumbers(L.size(), R.size()))
482 // Compare strings lexicographically only when it is necessary: only when
483 // strings are equal in size.
487 int FunctionComparator::cmpAttrs(const AttributeSet L,
488 const AttributeSet R) const {
489 if (int Res = cmpNumbers(L.getNumSlots(), R.getNumSlots()))
492 for (unsigned i = 0, e = L.getNumSlots(); i != e; ++i) {
493 AttributeSet::iterator LI = L.begin(i), LE = L.end(i), RI = R.begin(i),
495 for (; LI != LE && RI != RE; ++LI, ++RI) {
511 /// Constants comparison:
512 /// 1. Check whether type of L constant could be losslessly bitcasted to R
514 /// 2. Compare constant contents.
515 /// For more details see declaration comments.
516 int FunctionComparator::cmpConstants(const Constant *L, const Constant *R) {
518 Type *TyL = L->getType();
519 Type *TyR = R->getType();
521 // Check whether types are bitcastable. This part is just re-factored
522 // Type::canLosslesslyBitCastTo method, but instead of returning true/false,
523 // we also pack into result which type is "less" for us.
524 int TypesRes = cmpType(TyL, TyR);
526 // Types are different, but check whether we can bitcast them.
527 if (!TyL->isFirstClassType()) {
528 if (TyR->isFirstClassType())
530 // Neither TyL nor TyR are values of first class type. Return the result
531 // of comparing the types
534 if (!TyR->isFirstClassType()) {
535 if (TyL->isFirstClassType())
540 // Vector -> Vector conversions are always lossless if the two vector types
541 // have the same size, otherwise not.
542 unsigned TyLWidth = 0;
543 unsigned TyRWidth = 0;
545 if (const VectorType *VecTyL = dyn_cast<VectorType>(TyL))
546 TyLWidth = VecTyL->getBitWidth();
547 if (const VectorType *VecTyR = dyn_cast<VectorType>(TyR))
548 TyRWidth = VecTyR->getBitWidth();
550 if (TyLWidth != TyRWidth)
551 return cmpNumbers(TyLWidth, TyRWidth);
553 // Zero bit-width means neither TyL nor TyR are vectors.
555 PointerType *PTyL = dyn_cast<PointerType>(TyL);
556 PointerType *PTyR = dyn_cast<PointerType>(TyR);
558 unsigned AddrSpaceL = PTyL->getAddressSpace();
559 unsigned AddrSpaceR = PTyR->getAddressSpace();
560 if (int Res = cmpNumbers(AddrSpaceL, AddrSpaceR))
568 // TyL and TyR aren't vectors, nor pointers. We don't know how to
574 // OK, types are bitcastable, now check constant contents.
576 if (L->isNullValue() && R->isNullValue())
578 if (L->isNullValue() && !R->isNullValue())
580 if (!L->isNullValue() && R->isNullValue())
583 if (int Res = cmpNumbers(L->getValueID(), R->getValueID()))
586 switch (L->getValueID()) {
587 case Value::UndefValueVal: return TypesRes;
588 case Value::ConstantIntVal: {
589 const APInt &LInt = cast<ConstantInt>(L)->getValue();
590 const APInt &RInt = cast<ConstantInt>(R)->getValue();
591 return cmpAPInt(LInt, RInt);
593 case Value::ConstantFPVal: {
594 const APFloat &LAPF = cast<ConstantFP>(L)->getValueAPF();
595 const APFloat &RAPF = cast<ConstantFP>(R)->getValueAPF();
596 return cmpAPFloat(LAPF, RAPF);
598 case Value::ConstantArrayVal: {
599 const ConstantArray *LA = cast<ConstantArray>(L);
600 const ConstantArray *RA = cast<ConstantArray>(R);
601 uint64_t NumElementsL = cast<ArrayType>(TyL)->getNumElements();
602 uint64_t NumElementsR = cast<ArrayType>(TyR)->getNumElements();
603 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
605 for (uint64_t i = 0; i < NumElementsL; ++i) {
606 if (int Res = cmpConstants(cast<Constant>(LA->getOperand(i)),
607 cast<Constant>(RA->getOperand(i))))
612 case Value::ConstantStructVal: {
613 const ConstantStruct *LS = cast<ConstantStruct>(L);
614 const ConstantStruct *RS = cast<ConstantStruct>(R);
615 unsigned NumElementsL = cast<StructType>(TyL)->getNumElements();
616 unsigned NumElementsR = cast<StructType>(TyR)->getNumElements();
617 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
619 for (unsigned i = 0; i != NumElementsL; ++i) {
620 if (int Res = cmpConstants(cast<Constant>(LS->getOperand(i)),
621 cast<Constant>(RS->getOperand(i))))
626 case Value::ConstantVectorVal: {
627 const ConstantVector *LV = cast<ConstantVector>(L);
628 const ConstantVector *RV = cast<ConstantVector>(R);
629 unsigned NumElementsL = cast<VectorType>(TyL)->getNumElements();
630 unsigned NumElementsR = cast<VectorType>(TyR)->getNumElements();
631 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
633 for (uint64_t i = 0; i < NumElementsL; ++i) {
634 if (int Res = cmpConstants(cast<Constant>(LV->getOperand(i)),
635 cast<Constant>(RV->getOperand(i))))
640 case Value::ConstantExprVal: {
641 const ConstantExpr *LE = cast<ConstantExpr>(L);
642 const ConstantExpr *RE = cast<ConstantExpr>(R);
643 unsigned NumOperandsL = LE->getNumOperands();
644 unsigned NumOperandsR = RE->getNumOperands();
645 if (int Res = cmpNumbers(NumOperandsL, NumOperandsR))
647 for (unsigned i = 0; i < NumOperandsL; ++i) {
648 if (int Res = cmpConstants(cast<Constant>(LE->getOperand(i)),
649 cast<Constant>(RE->getOperand(i))))
654 case Value::FunctionVal:
655 case Value::GlobalVariableVal:
656 case Value::GlobalAliasVal:
657 default: // Unknown constant, cast L and R pointers to numbers and compare.
658 return cmpNumbers((uint64_t)L, (uint64_t)R);
662 /// cmpType - compares two types,
663 /// defines total ordering among the types set.
664 /// See method declaration comments for more details.
665 int FunctionComparator::cmpType(Type *TyL, Type *TyR) const {
667 PointerType *PTyL = dyn_cast<PointerType>(TyL);
668 PointerType *PTyR = dyn_cast<PointerType>(TyR);
671 if (PTyL && PTyL->getAddressSpace() == 0) TyL = DL->getIntPtrType(TyL);
672 if (PTyR && PTyR->getAddressSpace() == 0) TyR = DL->getIntPtrType(TyR);
678 if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
681 switch (TyL->getTypeID()) {
683 llvm_unreachable("Unknown type!");
684 // Fall through in Release mode.
685 case Type::IntegerTyID:
686 case Type::VectorTyID:
687 // TyL == TyR would have returned true earlier.
688 return cmpNumbers((uint64_t)TyL, (uint64_t)TyR);
691 case Type::FloatTyID:
692 case Type::DoubleTyID:
693 case Type::X86_FP80TyID:
694 case Type::FP128TyID:
695 case Type::PPC_FP128TyID:
696 case Type::LabelTyID:
697 case Type::MetadataTyID:
700 case Type::PointerTyID: {
701 assert(PTyL && PTyR && "Both types must be pointers here.");
702 return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
705 case Type::StructTyID: {
706 StructType *STyL = cast<StructType>(TyL);
707 StructType *STyR = cast<StructType>(TyR);
708 if (STyL->getNumElements() != STyR->getNumElements())
709 return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
711 if (STyL->isPacked() != STyR->isPacked())
712 return cmpNumbers(STyL->isPacked(), STyR->isPacked());
714 for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
715 if (int Res = cmpType(STyL->getElementType(i),
716 STyR->getElementType(i)))
722 case Type::FunctionTyID: {
723 FunctionType *FTyL = cast<FunctionType>(TyL);
724 FunctionType *FTyR = cast<FunctionType>(TyR);
725 if (FTyL->getNumParams() != FTyR->getNumParams())
726 return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
728 if (FTyL->isVarArg() != FTyR->isVarArg())
729 return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
731 if (int Res = cmpType(FTyL->getReturnType(), FTyR->getReturnType()))
734 for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
735 if (int Res = cmpType(FTyL->getParamType(i), FTyR->getParamType(i)))
741 case Type::ArrayTyID: {
742 ArrayType *ATyL = cast<ArrayType>(TyL);
743 ArrayType *ATyR = cast<ArrayType>(TyR);
744 if (ATyL->getNumElements() != ATyR->getNumElements())
745 return cmpNumbers(ATyL->getNumElements(), ATyR->getNumElements());
746 return cmpType(ATyL->getElementType(), ATyR->getElementType());
751 // Determine whether the two operations are the same except that pointer-to-A
752 // and pointer-to-B are equivalent. This should be kept in sync with
753 // Instruction::isSameOperationAs.
754 // Read method declaration comments for more details.
755 int FunctionComparator::cmpOperation(const Instruction *L,
756 const Instruction *R) const {
757 // Differences from Instruction::isSameOperationAs:
758 // * replace type comparison with calls to isEquivalentType.
759 // * we test for I->hasSameSubclassOptionalData (nuw/nsw/tail) at the top
760 // * because of the above, we don't test for the tail bit on calls later on
761 if (int Res = cmpNumbers(L->getOpcode(), R->getOpcode()))
764 if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
767 if (int Res = cmpType(L->getType(), R->getType()))
770 if (int Res = cmpNumbers(L->getRawSubclassOptionalData(),
771 R->getRawSubclassOptionalData()))
774 // We have two instructions of identical opcode and #operands. Check to see
775 // if all operands are the same type
776 for (unsigned i = 0, e = L->getNumOperands(); i != e; ++i) {
778 cmpType(L->getOperand(i)->getType(), R->getOperand(i)->getType()))
782 // Check special state that is a part of some instructions.
783 if (const LoadInst *LI = dyn_cast<LoadInst>(L)) {
784 if (int Res = cmpNumbers(LI->isVolatile(), cast<LoadInst>(R)->isVolatile()))
787 cmpNumbers(LI->getAlignment(), cast<LoadInst>(R)->getAlignment()))
790 cmpNumbers(LI->getOrdering(), cast<LoadInst>(R)->getOrdering()))
793 cmpNumbers(LI->getSynchScope(), cast<LoadInst>(R)->getSynchScope()))
795 return cmpNumbers((uint64_t)LI->getMetadata(LLVMContext::MD_range),
796 (uint64_t)cast<LoadInst>(R)->getMetadata(LLVMContext::MD_range));
798 if (const StoreInst *SI = dyn_cast<StoreInst>(L)) {
800 cmpNumbers(SI->isVolatile(), cast<StoreInst>(R)->isVolatile()))
803 cmpNumbers(SI->getAlignment(), cast<StoreInst>(R)->getAlignment()))
806 cmpNumbers(SI->getOrdering(), cast<StoreInst>(R)->getOrdering()))
808 return cmpNumbers(SI->getSynchScope(), cast<StoreInst>(R)->getSynchScope());
810 if (const CmpInst *CI = dyn_cast<CmpInst>(L))
811 return cmpNumbers(CI->getPredicate(), cast<CmpInst>(R)->getPredicate());
812 if (const CallInst *CI = dyn_cast<CallInst>(L)) {
813 if (int Res = cmpNumbers(CI->getCallingConv(),
814 cast<CallInst>(R)->getCallingConv()))
816 return cmpAttrs(CI->getAttributes(), cast<CallInst>(R)->getAttributes());
818 if (const InvokeInst *CI = dyn_cast<InvokeInst>(L)) {
819 if (int Res = cmpNumbers(CI->getCallingConv(),
820 cast<InvokeInst>(R)->getCallingConv()))
822 return cmpAttrs(CI->getAttributes(), cast<InvokeInst>(R)->getAttributes());
824 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(L)) {
825 ArrayRef<unsigned> LIndices = IVI->getIndices();
826 ArrayRef<unsigned> RIndices = cast<InsertValueInst>(R)->getIndices();
827 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
829 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
830 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
834 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(L)) {
835 ArrayRef<unsigned> LIndices = EVI->getIndices();
836 ArrayRef<unsigned> RIndices = cast<ExtractValueInst>(R)->getIndices();
837 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
839 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
840 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
844 if (const FenceInst *FI = dyn_cast<FenceInst>(L)) {
846 cmpNumbers(FI->getOrdering(), cast<FenceInst>(R)->getOrdering()))
848 return cmpNumbers(FI->getSynchScope(), cast<FenceInst>(R)->getSynchScope());
851 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(L)) {
852 if (int Res = cmpNumbers(CXI->isVolatile(),
853 cast<AtomicCmpXchgInst>(R)->isVolatile()))
855 if (int Res = cmpNumbers(CXI->isWeak(),
856 cast<AtomicCmpXchgInst>(R)->isWeak()))
858 if (int Res = cmpNumbers(CXI->getSuccessOrdering(),
859 cast<AtomicCmpXchgInst>(R)->getSuccessOrdering()))
861 if (int Res = cmpNumbers(CXI->getFailureOrdering(),
862 cast<AtomicCmpXchgInst>(R)->getFailureOrdering()))
864 return cmpNumbers(CXI->getSynchScope(),
865 cast<AtomicCmpXchgInst>(R)->getSynchScope());
867 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(L)) {
868 if (int Res = cmpNumbers(RMWI->getOperation(),
869 cast<AtomicRMWInst>(R)->getOperation()))
871 if (int Res = cmpNumbers(RMWI->isVolatile(),
872 cast<AtomicRMWInst>(R)->isVolatile()))
874 if (int Res = cmpNumbers(RMWI->getOrdering(),
875 cast<AtomicRMWInst>(R)->getOrdering()))
877 return cmpNumbers(RMWI->getSynchScope(),
878 cast<AtomicRMWInst>(R)->getSynchScope());
883 // Determine whether two GEP operations perform the same underlying arithmetic.
884 // Read method declaration comments for more details.
885 int FunctionComparator::cmpGEP(const GEPOperator *GEPL,
886 const GEPOperator *GEPR) {
888 unsigned int ASL = GEPL->getPointerAddressSpace();
889 unsigned int ASR = GEPR->getPointerAddressSpace();
891 if (int Res = cmpNumbers(ASL, ASR))
894 // When we have target data, we can reduce the GEP down to the value in bytes
895 // added to the address.
897 unsigned BitWidth = DL->getPointerSizeInBits(ASL);
898 APInt OffsetL(BitWidth, 0), OffsetR(BitWidth, 0);
899 if (GEPL->accumulateConstantOffset(*DL, OffsetL) &&
900 GEPR->accumulateConstantOffset(*DL, OffsetR))
901 return cmpAPInt(OffsetL, OffsetR);
904 if (int Res = cmpNumbers((uint64_t)GEPL->getPointerOperand()->getType(),
905 (uint64_t)GEPR->getPointerOperand()->getType()))
908 if (int Res = cmpNumbers(GEPL->getNumOperands(), GEPR->getNumOperands()))
911 for (unsigned i = 0, e = GEPL->getNumOperands(); i != e; ++i) {
912 if (int Res = cmpValues(GEPL->getOperand(i), GEPR->getOperand(i)))
919 /// Compare two values used by the two functions under pair-wise comparison. If
920 /// this is the first time the values are seen, they're added to the mapping so
921 /// that we will detect mismatches on next use.
922 /// See comments in declaration for more details.
923 int FunctionComparator::cmpValues(const Value *L, const Value *R) {
924 // Catch self-reference case.
936 const Constant *ConstL = dyn_cast<Constant>(L);
937 const Constant *ConstR = dyn_cast<Constant>(R);
938 if (ConstL && ConstR) {
941 return cmpConstants(ConstL, ConstR);
949 const InlineAsm *InlineAsmL = dyn_cast<InlineAsm>(L);
950 const InlineAsm *InlineAsmR = dyn_cast<InlineAsm>(R);
952 if (InlineAsmL && InlineAsmR)
953 return cmpNumbers((uint64_t)L, (uint64_t)R);
959 auto LeftSN = sn_mapL.insert(std::make_pair(L, sn_mapL.size())),
960 RightSN = sn_mapR.insert(std::make_pair(R, sn_mapR.size()));
962 return cmpNumbers(LeftSN.first->second, RightSN.first->second);
964 // Test whether two basic blocks have equivalent behaviour.
965 bool FunctionComparator::compare(const BasicBlock *BB1, const BasicBlock *BB2) {
966 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end();
967 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end();
970 if (!enumerate(F1I, F2I))
973 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) {
974 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I);
978 if (!enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand()))
981 if (!isEquivalentGEP(GEP1, GEP2))
984 if (!isEquivalentOperation(F1I, F2I))
987 assert(F1I->getNumOperands() == F2I->getNumOperands());
988 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) {
989 Value *OpF1 = F1I->getOperand(i);
990 Value *OpF2 = F2I->getOperand(i);
992 if (!enumerate(OpF1, OpF2))
995 if (OpF1->getValueID() != OpF2->getValueID() ||
996 !isEquivalentType(OpF1->getType(), OpF2->getType()))
1002 } while (F1I != F1E && F2I != F2E);
1004 return F1I == F1E && F2I == F2E;
1007 // Test whether the two functions have equivalent behaviour.
1008 bool FunctionComparator::compare() {
1009 // We need to recheck everything, but check the things that weren't included
1010 // in the hash first.
1015 if (F1->getAttributes() != F2->getAttributes())
1018 if (F1->hasGC() != F2->hasGC())
1021 if (F1->hasGC() && F1->getGC() != F2->getGC())
1024 if (F1->hasSection() != F2->hasSection())
1027 if (F1->hasSection() && F1->getSection() != F2->getSection())
1030 if (F1->isVarArg() != F2->isVarArg())
1033 // TODO: if it's internal and only used in direct calls, we could handle this
1035 if (F1->getCallingConv() != F2->getCallingConv())
1038 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType()))
1041 assert(F1->arg_size() == F2->arg_size() &&
1042 "Identically typed functions have different numbers of args!");
1044 // Visit the arguments so that they get enumerated in the order they're
1046 for (Function::const_arg_iterator f1i = F1->arg_begin(),
1047 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) {
1048 if (!enumerate(f1i, f2i))
1049 llvm_unreachable("Arguments repeat!");
1052 // We do a CFG-ordered walk since the actual ordering of the blocks in the
1053 // linked list is immaterial. Our walk starts at the entry block for both
1054 // functions, then takes each block from each terminator in order. As an
1055 // artifact, this also means that unreachable blocks are ignored.
1056 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs;
1057 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
1059 F1BBs.push_back(&F1->getEntryBlock());
1060 F2BBs.push_back(&F2->getEntryBlock());
1062 VisitedBBs.insert(F1BBs[0]);
1063 while (!F1BBs.empty()) {
1064 const BasicBlock *F1BB = F1BBs.pop_back_val();
1065 const BasicBlock *F2BB = F2BBs.pop_back_val();
1067 if (!enumerate(F1BB, F2BB) || !compare(F1BB, F2BB))
1070 const TerminatorInst *F1TI = F1BB->getTerminator();
1071 const TerminatorInst *F2TI = F2BB->getTerminator();
1073 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors());
1074 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) {
1075 if (!VisitedBBs.insert(F1TI->getSuccessor(i)))
1078 F1BBs.push_back(F1TI->getSuccessor(i));
1079 F2BBs.push_back(F2TI->getSuccessor(i));
1087 /// MergeFunctions finds functions which will generate identical machine code,
1088 /// by considering all pointer types to be equivalent. Once identified,
1089 /// MergeFunctions will fold them by replacing a call to one to a call to a
1090 /// bitcast of the other.
1092 class MergeFunctions : public ModulePass {
1096 : ModulePass(ID), HasGlobalAliases(false) {
1097 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
1100 bool runOnModule(Module &M) override;
1103 typedef DenseSet<ComparableFunction> FnSetType;
1105 /// A work queue of functions that may have been modified and should be
1107 std::vector<WeakVH> Deferred;
1109 /// Insert a ComparableFunction into the FnSet, or merge it away if it's
1110 /// equal to one that's already present.
1111 bool insert(ComparableFunction &NewF);
1113 /// Remove a Function from the FnSet and queue it up for a second sweep of
1115 void remove(Function *F);
1117 /// Find the functions that use this Value and remove them from FnSet and
1118 /// queue the functions.
1119 void removeUsers(Value *V);
1121 /// Replace all direct calls of Old with calls of New. Will bitcast New if
1122 /// necessary to make types match.
1123 void replaceDirectCallers(Function *Old, Function *New);
1125 /// Merge two equivalent functions. Upon completion, G may be deleted, or may
1126 /// be converted into a thunk. In either case, it should never be visited
1128 void mergeTwoFunctions(Function *F, Function *G);
1130 /// Replace G with a thunk or an alias to F. Deletes G.
1131 void writeThunkOrAlias(Function *F, Function *G);
1133 /// Replace G with a simple tail call to bitcast(F). Also replace direct uses
1134 /// of G with bitcast(F). Deletes G.
1135 void writeThunk(Function *F, Function *G);
1137 /// Replace G with an alias to F. Deletes G.
1138 void writeAlias(Function *F, Function *G);
1140 /// The set of all distinct functions. Use the insert() and remove() methods
1144 /// DataLayout for more accurate GEP comparisons. May be NULL.
1145 const DataLayout *DL;
1147 /// Whether or not the target supports global aliases.
1148 bool HasGlobalAliases;
1151 } // end anonymous namespace
1153 char MergeFunctions::ID = 0;
1154 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
1156 ModulePass *llvm::createMergeFunctionsPass() {
1157 return new MergeFunctions();
1160 bool MergeFunctions::runOnModule(Module &M) {
1161 bool Changed = false;
1162 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1163 DL = DLP ? &DLP->getDataLayout() : nullptr;
1165 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
1166 if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage())
1167 Deferred.push_back(WeakVH(I));
1169 FnSet.resize(Deferred.size());
1172 std::vector<WeakVH> Worklist;
1173 Deferred.swap(Worklist);
1175 DEBUG(dbgs() << "size of module: " << M.size() << '\n');
1176 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
1178 // Insert only strong functions and merge them. Strong function merging
1179 // always deletes one of them.
1180 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
1181 E = Worklist.end(); I != E; ++I) {
1183 Function *F = cast<Function>(*I);
1184 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
1185 !F->mayBeOverridden()) {
1186 ComparableFunction CF = ComparableFunction(F, DL);
1187 Changed |= insert(CF);
1191 // Insert only weak functions and merge them. By doing these second we
1192 // create thunks to the strong function when possible. When two weak
1193 // functions are identical, we create a new strong function with two weak
1194 // weak thunks to it which are identical but not mergable.
1195 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
1196 E = Worklist.end(); I != E; ++I) {
1198 Function *F = cast<Function>(*I);
1199 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
1200 F->mayBeOverridden()) {
1201 ComparableFunction CF = ComparableFunction(F, DL);
1202 Changed |= insert(CF);
1205 DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n');
1206 } while (!Deferred.empty());
1213 bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS,
1214 const ComparableFunction &RHS) {
1215 if (LHS.getFunc() == RHS.getFunc() &&
1216 LHS.getHash() == RHS.getHash())
1218 if (!LHS.getFunc() || !RHS.getFunc())
1221 // One of these is a special "underlying pointer comparison only" object.
1222 if (LHS.getDataLayout() == ComparableFunction::LookupOnly ||
1223 RHS.getDataLayout() == ComparableFunction::LookupOnly)
1226 assert(LHS.getDataLayout() == RHS.getDataLayout() &&
1227 "Comparing functions for different targets");
1229 return FunctionComparator(LHS.getDataLayout(), LHS.getFunc(),
1230 RHS.getFunc()).compare();
1233 // Replace direct callers of Old with New.
1234 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
1235 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
1236 for (auto UI = Old->use_begin(), UE = Old->use_end(); UI != UE;) {
1239 CallSite CS(U->getUser());
1240 if (CS && CS.isCallee(U)) {
1241 remove(CS.getInstruction()->getParent()->getParent());
1247 // Replace G with an alias to F if possible, or else a thunk to F. Deletes G.
1248 void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) {
1249 if (HasGlobalAliases && G->hasUnnamedAddr()) {
1250 if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
1251 G->hasWeakLinkage()) {
1260 // Helper for writeThunk,
1261 // Selects proper bitcast operation,
1262 // but a bit simpler then CastInst::getCastOpcode.
1263 static Value *createCast(IRBuilder<false> &Builder, Value *V, Type *DestTy) {
1264 Type *SrcTy = V->getType();
1265 if (SrcTy->isStructTy()) {
1266 assert(DestTy->isStructTy());
1267 assert(SrcTy->getStructNumElements() == DestTy->getStructNumElements());
1268 Value *Result = UndefValue::get(DestTy);
1269 for (unsigned int I = 0, E = SrcTy->getStructNumElements(); I < E; ++I) {
1270 Value *Element = createCast(
1271 Builder, Builder.CreateExtractValue(V, ArrayRef<unsigned int>(I)),
1272 DestTy->getStructElementType(I));
1275 Builder.CreateInsertValue(Result, Element, ArrayRef<unsigned int>(I));
1279 assert(!DestTy->isStructTy());
1280 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
1281 return Builder.CreateIntToPtr(V, DestTy);
1282 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
1283 return Builder.CreatePtrToInt(V, DestTy);
1285 return Builder.CreateBitCast(V, DestTy);
1288 // Replace G with a simple tail call to bitcast(F). Also replace direct uses
1289 // of G with bitcast(F). Deletes G.
1290 void MergeFunctions::writeThunk(Function *F, Function *G) {
1291 if (!G->mayBeOverridden()) {
1292 // Redirect direct callers of G to F.
1293 replaceDirectCallers(G, F);
1296 // If G was internal then we may have replaced all uses of G with F. If so,
1297 // stop here and delete G. There's no need for a thunk.
1298 if (G->hasLocalLinkage() && G->use_empty()) {
1299 G->eraseFromParent();
1303 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
1305 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
1306 IRBuilder<false> Builder(BB);
1308 SmallVector<Value *, 16> Args;
1310 FunctionType *FFTy = F->getFunctionType();
1311 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
1313 Args.push_back(createCast(Builder, (Value*)AI, FFTy->getParamType(i)));
1317 CallInst *CI = Builder.CreateCall(F, Args);
1319 CI->setCallingConv(F->getCallingConv());
1320 if (NewG->getReturnType()->isVoidTy()) {
1321 Builder.CreateRetVoid();
1323 Builder.CreateRet(createCast(Builder, CI, NewG->getReturnType()));
1326 NewG->copyAttributesFrom(G);
1329 G->replaceAllUsesWith(NewG);
1330 G->eraseFromParent();
1332 DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n');
1336 // Replace G with an alias to F and delete G.
1337 void MergeFunctions::writeAlias(Function *F, Function *G) {
1338 PointerType *PTy = G->getType();
1339 auto *GA = GlobalAlias::create(PTy->getElementType(), PTy->getAddressSpace(),
1340 G->getLinkage(), "", F);
1341 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
1343 GA->setVisibility(G->getVisibility());
1345 G->replaceAllUsesWith(GA);
1346 G->eraseFromParent();
1348 DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n');
1349 ++NumAliasesWritten;
1352 // Merge two equivalent functions. Upon completion, Function G is deleted.
1353 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) {
1354 if (F->mayBeOverridden()) {
1355 assert(G->mayBeOverridden());
1357 if (HasGlobalAliases) {
1358 // Make them both thunks to the same internal function.
1359 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
1361 H->copyAttributesFrom(F);
1364 F->replaceAllUsesWith(H);
1366 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
1371 F->setAlignment(MaxAlignment);
1372 F->setLinkage(GlobalValue::PrivateLinkage);
1374 // We can't merge them. Instead, pick one and update all direct callers
1375 // to call it and hope that we improve the instruction cache hit rate.
1376 replaceDirectCallers(G, F);
1381 writeThunkOrAlias(F, G);
1384 ++NumFunctionsMerged;
1387 // Insert a ComparableFunction into the FnSet, or merge it away if equal to one
1388 // that was already inserted.
1389 bool MergeFunctions::insert(ComparableFunction &NewF) {
1390 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF);
1391 if (Result.second) {
1392 DEBUG(dbgs() << "Inserting as unique: " << NewF.getFunc()->getName() << '\n');
1396 const ComparableFunction &OldF = *Result.first;
1398 // Don't merge tiny functions, since it can just end up making the function
1400 // FIXME: Should still merge them if they are unnamed_addr and produce an
1402 if (NewF.getFunc()->size() == 1) {
1403 if (NewF.getFunc()->front().size() <= 2) {
1404 DEBUG(dbgs() << NewF.getFunc()->getName()
1405 << " is to small to bother merging\n");
1410 // Never thunk a strong function to a weak function.
1411 assert(!OldF.getFunc()->mayBeOverridden() ||
1412 NewF.getFunc()->mayBeOverridden());
1414 DEBUG(dbgs() << " " << OldF.getFunc()->getName() << " == "
1415 << NewF.getFunc()->getName() << '\n');
1417 Function *DeleteF = NewF.getFunc();
1419 mergeTwoFunctions(OldF.getFunc(), DeleteF);
1423 // Remove a function from FnSet. If it was already in FnSet, add it to Deferred
1424 // so that we'll look at it in the next round.
1425 void MergeFunctions::remove(Function *F) {
1426 // We need to make sure we remove F, not a function "equal" to F per the
1427 // function equality comparator.
1429 // The special "lookup only" ComparableFunction bypasses the expensive
1430 // function comparison in favour of a pointer comparison on the underlying
1432 ComparableFunction CF = ComparableFunction(F, ComparableFunction::LookupOnly);
1433 if (FnSet.erase(CF)) {
1434 DEBUG(dbgs() << "Removed " << F->getName() << " from set and deferred it.\n");
1435 Deferred.push_back(F);
1439 // For each instruction used by the value, remove() the function that contains
1440 // the instruction. This should happen right before a call to RAUW.
1441 void MergeFunctions::removeUsers(Value *V) {
1442 std::vector<Value *> Worklist;
1443 Worklist.push_back(V);
1444 while (!Worklist.empty()) {
1445 Value *V = Worklist.back();
1446 Worklist.pop_back();
1448 for (User *U : V->users()) {
1449 if (Instruction *I = dyn_cast<Instruction>(U)) {
1450 remove(I->getParent()->getParent());
1451 } else if (isa<GlobalValue>(U)) {
1453 } else if (Constant *C = dyn_cast<Constant>(U)) {
1454 for (User *UU : C->users())
1455 Worklist.push_back(UU);