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 bool isEquivalentOperation(const Instruction *I1,
311 const Instruction *I2) const;
313 /// Compare two GEPs for equivalent pointer arithmetic.
314 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2);
315 bool isEquivalentGEP(const GetElementPtrInst *GEP1,
316 const GetElementPtrInst *GEP2) {
317 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2));
320 /// cmpType - compares two types,
321 /// defines total ordering among the types set.
324 /// 0 if types are equal,
325 /// -1 if Left is less than Right,
326 /// +1 if Left is greater than Right.
329 /// Comparison is broken onto stages. Like in lexicographical comparison
330 /// stage coming first has higher priority.
331 /// On each explanation stage keep in mind total ordering properties.
333 /// 0. Before comparison we coerce pointer types of 0 address space to
335 /// We also don't bother with same type at left and right, so
336 /// just return 0 in this case.
338 /// 1. If types are of different kind (different type IDs).
339 /// Return result of type IDs comparison, treating them as numbers.
340 /// 2. If types are vectors or integers, compare Type* values as numbers.
341 /// 3. Types has same ID, so check whether they belongs to the next group:
350 /// If so - return 0, yes - we can treat these types as equal only because
351 /// their IDs are same.
352 /// 4. If Left and Right are pointers, return result of address space
353 /// comparison (numbers comparison). We can treat pointer types of same
354 /// address space as equal.
355 /// 5. If types are complex.
356 /// Then both Left and Right are to be expanded and their element types will
357 /// be checked with the same way. If we get Res != 0 on some stage, return it.
358 /// Otherwise return 0.
359 /// 6. For all other cases put llvm_unreachable.
360 int cmpType(Type *TyL, Type *TyR) const;
362 bool isEquivalentType(Type *Ty1, Type *Ty2) const {
363 return cmpType(Ty1, Ty2) == 0;
366 int cmpNumbers(uint64_t L, uint64_t R) const;
368 int cmpAPInt(const APInt &L, const APInt &R) const;
369 int cmpAPFloat(const APFloat &L, const APFloat &R) const;
371 // The two functions undergoing comparison.
372 const Function *F1, *F2;
374 const DataLayout *DL;
376 /// Assign serial numbers to values from left function, and values from
379 /// Being comparing functions we need to compare values we meet at left and
381 /// Its easy to sort things out for external values. It just should be
382 /// the same value at left and right.
383 /// But for local values (those were introduced inside function body)
384 /// we have to ensure they were introduced at exactly the same place,
385 /// and plays the same role.
386 /// Let's assign serial number to each value when we meet it first time.
387 /// Values that were met at same place will be with same serial numbers.
388 /// In this case it would be good to explain few points about values assigned
389 /// to BBs and other ways of implementation (see below).
391 /// 1. Safety of BB reordering.
392 /// It's safe to change the order of BasicBlocks in function.
393 /// Relationship with other functions and serial numbering will not be
394 /// changed in this case.
395 /// As follows from FunctionComparator::compare(), we do CFG walk: we start
396 /// from the entry, and then take each terminator. So it doesn't matter how in
397 /// fact BBs are ordered in function. And since cmpValues are called during
398 /// this walk, the numbering depends only on how BBs located inside the CFG.
399 /// So the answer is - yes. We will get the same numbering.
401 /// 2. Impossibility to use dominance properties of values.
402 /// If we compare two instruction operands: first is usage of local
403 /// variable AL from function FL, and second is usage of local variable AR
404 /// from FR, we could compare their origins and check whether they are
405 /// defined at the same place.
406 /// But, we are still not able to compare operands of PHI nodes, since those
407 /// could be operands from further BBs we didn't scan yet.
408 /// So it's impossible to use dominance properties in general.
409 DenseMap<const Value*, int> sn_mapL, sn_mapR;
414 int FunctionComparator::cmpNumbers(uint64_t L, uint64_t R) const {
415 if (L < R) return -1;
420 int FunctionComparator::cmpAPInt(const APInt &L, const APInt &R) const {
421 if (int Res = cmpNumbers(L.getBitWidth(), R.getBitWidth()))
423 if (L.ugt(R)) return 1;
424 if (R.ugt(L)) return -1;
428 int FunctionComparator::cmpAPFloat(const APFloat &L, const APFloat &R) const {
429 if (int Res = cmpNumbers((uint64_t)&L.getSemantics(),
430 (uint64_t)&R.getSemantics()))
432 return cmpAPInt(L.bitcastToAPInt(), R.bitcastToAPInt());
435 /// Constants comparison:
436 /// 1. Check whether type of L constant could be losslessly bitcasted to R
438 /// 2. Compare constant contents.
439 /// For more details see declaration comments.
440 int FunctionComparator::cmpConstants(const Constant *L, const Constant *R) {
442 Type *TyL = L->getType();
443 Type *TyR = R->getType();
445 // Check whether types are bitcastable. This part is just re-factored
446 // Type::canLosslesslyBitCastTo method, but instead of returning true/false,
447 // we also pack into result which type is "less" for us.
448 int TypesRes = cmpType(TyL, TyR);
450 // Types are different, but check whether we can bitcast them.
451 if (!TyL->isFirstClassType()) {
452 if (TyR->isFirstClassType())
454 // Neither TyL nor TyR are values of first class type. Return the result
455 // of comparing the types
458 if (!TyR->isFirstClassType()) {
459 if (TyL->isFirstClassType())
464 // Vector -> Vector conversions are always lossless if the two vector types
465 // have the same size, otherwise not.
466 unsigned TyLWidth = 0;
467 unsigned TyRWidth = 0;
469 if (const VectorType *VecTyL = dyn_cast<VectorType>(TyL))
470 TyLWidth = VecTyL->getBitWidth();
471 if (const VectorType *VecTyR = dyn_cast<VectorType>(TyR))
472 TyRWidth = VecTyR->getBitWidth();
474 if (TyLWidth != TyRWidth)
475 return cmpNumbers(TyLWidth, TyRWidth);
477 // Zero bit-width means neither TyL nor TyR are vectors.
479 PointerType *PTyL = dyn_cast<PointerType>(TyL);
480 PointerType *PTyR = dyn_cast<PointerType>(TyR);
482 unsigned AddrSpaceL = PTyL->getAddressSpace();
483 unsigned AddrSpaceR = PTyR->getAddressSpace();
484 if (int Res = cmpNumbers(AddrSpaceL, AddrSpaceR))
492 // TyL and TyR aren't vectors, nor pointers. We don't know how to
498 // OK, types are bitcastable, now check constant contents.
500 if (L->isNullValue() && R->isNullValue())
502 if (L->isNullValue() && !R->isNullValue())
504 if (!L->isNullValue() && R->isNullValue())
507 if (int Res = cmpNumbers(L->getValueID(), R->getValueID()))
510 switch (L->getValueID()) {
511 case Value::UndefValueVal: return TypesRes;
512 case Value::ConstantIntVal: {
513 const APInt &LInt = cast<ConstantInt>(L)->getValue();
514 const APInt &RInt = cast<ConstantInt>(R)->getValue();
515 return cmpAPInt(LInt, RInt);
517 case Value::ConstantFPVal: {
518 const APFloat &LAPF = cast<ConstantFP>(L)->getValueAPF();
519 const APFloat &RAPF = cast<ConstantFP>(R)->getValueAPF();
520 return cmpAPFloat(LAPF, RAPF);
522 case Value::ConstantArrayVal: {
523 const ConstantArray *LA = cast<ConstantArray>(L);
524 const ConstantArray *RA = cast<ConstantArray>(R);
525 uint64_t NumElementsL = cast<ArrayType>(TyL)->getNumElements();
526 uint64_t NumElementsR = cast<ArrayType>(TyR)->getNumElements();
527 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
529 for (uint64_t i = 0; i < NumElementsL; ++i) {
530 if (int Res = cmpConstants(cast<Constant>(LA->getOperand(i)),
531 cast<Constant>(RA->getOperand(i))))
536 case Value::ConstantStructVal: {
537 const ConstantStruct *LS = cast<ConstantStruct>(L);
538 const ConstantStruct *RS = cast<ConstantStruct>(R);
539 unsigned NumElementsL = cast<StructType>(TyL)->getNumElements();
540 unsigned NumElementsR = cast<StructType>(TyR)->getNumElements();
541 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
543 for (unsigned i = 0; i != NumElementsL; ++i) {
544 if (int Res = cmpConstants(cast<Constant>(LS->getOperand(i)),
545 cast<Constant>(RS->getOperand(i))))
550 case Value::ConstantVectorVal: {
551 const ConstantVector *LV = cast<ConstantVector>(L);
552 const ConstantVector *RV = cast<ConstantVector>(R);
553 unsigned NumElementsL = cast<VectorType>(TyL)->getNumElements();
554 unsigned NumElementsR = cast<VectorType>(TyR)->getNumElements();
555 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
557 for (uint64_t i = 0; i < NumElementsL; ++i) {
558 if (int Res = cmpConstants(cast<Constant>(LV->getOperand(i)),
559 cast<Constant>(RV->getOperand(i))))
564 case Value::ConstantExprVal: {
565 const ConstantExpr *LE = cast<ConstantExpr>(L);
566 const ConstantExpr *RE = cast<ConstantExpr>(R);
567 unsigned NumOperandsL = LE->getNumOperands();
568 unsigned NumOperandsR = RE->getNumOperands();
569 if (int Res = cmpNumbers(NumOperandsL, NumOperandsR))
571 for (unsigned i = 0; i < NumOperandsL; ++i) {
572 if (int Res = cmpConstants(cast<Constant>(LE->getOperand(i)),
573 cast<Constant>(RE->getOperand(i))))
578 case Value::FunctionVal:
579 case Value::GlobalVariableVal:
580 case Value::GlobalAliasVal:
581 default: // Unknown constant, cast L and R pointers to numbers and compare.
582 return cmpNumbers((uint64_t)L, (uint64_t)R);
586 /// cmpType - compares two types,
587 /// defines total ordering among the types set.
588 /// See method declaration comments for more details.
589 int FunctionComparator::cmpType(Type *TyL, Type *TyR) const {
591 PointerType *PTyL = dyn_cast<PointerType>(TyL);
592 PointerType *PTyR = dyn_cast<PointerType>(TyR);
595 if (PTyL && PTyL->getAddressSpace() == 0) TyL = DL->getIntPtrType(TyL);
596 if (PTyR && PTyR->getAddressSpace() == 0) TyR = DL->getIntPtrType(TyR);
602 if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
605 switch (TyL->getTypeID()) {
607 llvm_unreachable("Unknown type!");
608 // Fall through in Release mode.
609 case Type::IntegerTyID:
610 case Type::VectorTyID:
611 // TyL == TyR would have returned true earlier.
612 return cmpNumbers((uint64_t)TyL, (uint64_t)TyR);
615 case Type::FloatTyID:
616 case Type::DoubleTyID:
617 case Type::X86_FP80TyID:
618 case Type::FP128TyID:
619 case Type::PPC_FP128TyID:
620 case Type::LabelTyID:
621 case Type::MetadataTyID:
624 case Type::PointerTyID: {
625 assert(PTyL && PTyR && "Both types must be pointers here.");
626 return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
629 case Type::StructTyID: {
630 StructType *STyL = cast<StructType>(TyL);
631 StructType *STyR = cast<StructType>(TyR);
632 if (STyL->getNumElements() != STyR->getNumElements())
633 return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
635 if (STyL->isPacked() != STyR->isPacked())
636 return cmpNumbers(STyL->isPacked(), STyR->isPacked());
638 for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
639 if (int Res = cmpType(STyL->getElementType(i),
640 STyR->getElementType(i)))
646 case Type::FunctionTyID: {
647 FunctionType *FTyL = cast<FunctionType>(TyL);
648 FunctionType *FTyR = cast<FunctionType>(TyR);
649 if (FTyL->getNumParams() != FTyR->getNumParams())
650 return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
652 if (FTyL->isVarArg() != FTyR->isVarArg())
653 return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
655 if (int Res = cmpType(FTyL->getReturnType(), FTyR->getReturnType()))
658 for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
659 if (int Res = cmpType(FTyL->getParamType(i), FTyR->getParamType(i)))
665 case Type::ArrayTyID: {
666 ArrayType *ATyL = cast<ArrayType>(TyL);
667 ArrayType *ATyR = cast<ArrayType>(TyR);
668 if (ATyL->getNumElements() != ATyR->getNumElements())
669 return cmpNumbers(ATyL->getNumElements(), ATyR->getNumElements());
670 return cmpType(ATyL->getElementType(), ATyR->getElementType());
675 // Determine whether the two operations are the same except that pointer-to-A
676 // and pointer-to-B are equivalent. This should be kept in sync with
677 // Instruction::isSameOperationAs.
678 bool FunctionComparator::isEquivalentOperation(const Instruction *I1,
679 const Instruction *I2) const {
680 // Differences from Instruction::isSameOperationAs:
681 // * replace type comparison with calls to isEquivalentType.
682 // * we test for I->hasSameSubclassOptionalData (nuw/nsw/tail) at the top
683 // * because of the above, we don't test for the tail bit on calls later on
684 if (I1->getOpcode() != I2->getOpcode() ||
685 I1->getNumOperands() != I2->getNumOperands() ||
686 !isEquivalentType(I1->getType(), I2->getType()) ||
687 !I1->hasSameSubclassOptionalData(I2))
690 // We have two instructions of identical opcode and #operands. Check to see
691 // if all operands are the same type
692 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i)
693 if (!isEquivalentType(I1->getOperand(i)->getType(),
694 I2->getOperand(i)->getType()))
697 // Check special state that is a part of some instructions.
698 if (const LoadInst *LI = dyn_cast<LoadInst>(I1))
699 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() &&
700 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment() &&
701 LI->getOrdering() == cast<LoadInst>(I2)->getOrdering() &&
702 LI->getSynchScope() == cast<LoadInst>(I2)->getSynchScope();
703 if (const StoreInst *SI = dyn_cast<StoreInst>(I1))
704 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() &&
705 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment() &&
706 SI->getOrdering() == cast<StoreInst>(I2)->getOrdering() &&
707 SI->getSynchScope() == cast<StoreInst>(I2)->getSynchScope();
708 if (const CmpInst *CI = dyn_cast<CmpInst>(I1))
709 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate();
710 if (const CallInst *CI = dyn_cast<CallInst>(I1))
711 return CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() &&
712 CI->getAttributes() == cast<CallInst>(I2)->getAttributes();
713 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1))
714 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() &&
715 CI->getAttributes() == cast<InvokeInst>(I2)->getAttributes();
716 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1))
717 return IVI->getIndices() == cast<InsertValueInst>(I2)->getIndices();
718 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1))
719 return EVI->getIndices() == cast<ExtractValueInst>(I2)->getIndices();
720 if (const FenceInst *FI = dyn_cast<FenceInst>(I1))
721 return FI->getOrdering() == cast<FenceInst>(I2)->getOrdering() &&
722 FI->getSynchScope() == cast<FenceInst>(I2)->getSynchScope();
723 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I1))
724 return CXI->isVolatile() == cast<AtomicCmpXchgInst>(I2)->isVolatile() &&
725 CXI->getSuccessOrdering() ==
726 cast<AtomicCmpXchgInst>(I2)->getSuccessOrdering() &&
727 CXI->getFailureOrdering() ==
728 cast<AtomicCmpXchgInst>(I2)->getFailureOrdering() &&
729 CXI->getSynchScope() == cast<AtomicCmpXchgInst>(I2)->getSynchScope();
730 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I1))
731 return RMWI->getOperation() == cast<AtomicRMWInst>(I2)->getOperation() &&
732 RMWI->isVolatile() == cast<AtomicRMWInst>(I2)->isVolatile() &&
733 RMWI->getOrdering() == cast<AtomicRMWInst>(I2)->getOrdering() &&
734 RMWI->getSynchScope() == cast<AtomicRMWInst>(I2)->getSynchScope();
739 // Determine whether two GEP operations perform the same underlying arithmetic.
740 bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1,
741 const GEPOperator *GEP2) {
742 unsigned AS = GEP1->getPointerAddressSpace();
743 if (AS != GEP2->getPointerAddressSpace())
747 // When we have target data, we can reduce the GEP down to the value in bytes
748 // added to the address.
749 unsigned BitWidth = DL ? DL->getPointerSizeInBits(AS) : 1;
750 APInt Offset1(BitWidth, 0), Offset2(BitWidth, 0);
751 if (GEP1->accumulateConstantOffset(*DL, Offset1) &&
752 GEP2->accumulateConstantOffset(*DL, Offset2)) {
753 return Offset1 == Offset2;
757 if (GEP1->getPointerOperand()->getType() !=
758 GEP2->getPointerOperand()->getType())
761 if (GEP1->getNumOperands() != GEP2->getNumOperands())
764 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) {
765 if (!enumerate(GEP1->getOperand(i), GEP2->getOperand(i)))
772 /// Compare two values used by the two functions under pair-wise comparison. If
773 /// this is the first time the values are seen, they're added to the mapping so
774 /// that we will detect mismatches on next use.
775 /// See comments in declaration for more details.
776 int FunctionComparator::cmpValues(const Value *L, const Value *R) {
777 // Catch self-reference case.
789 const Constant *ConstL = dyn_cast<Constant>(L);
790 const Constant *ConstR = dyn_cast<Constant>(R);
791 if (ConstL && ConstR) {
794 return cmpConstants(ConstL, ConstR);
802 const InlineAsm *InlineAsmL = dyn_cast<InlineAsm>(L);
803 const InlineAsm *InlineAsmR = dyn_cast<InlineAsm>(R);
805 if (InlineAsmL && InlineAsmR)
806 return cmpNumbers((uint64_t)L, (uint64_t)R);
812 auto LeftSN = sn_mapL.insert(std::make_pair(L, sn_mapL.size())),
813 RightSN = sn_mapR.insert(std::make_pair(R, sn_mapR.size()));
815 return cmpNumbers(LeftSN.first->second, RightSN.first->second);
817 // Test whether two basic blocks have equivalent behaviour.
818 bool FunctionComparator::compare(const BasicBlock *BB1, const BasicBlock *BB2) {
819 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end();
820 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end();
823 if (!enumerate(F1I, F2I))
826 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) {
827 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I);
831 if (!enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand()))
834 if (!isEquivalentGEP(GEP1, GEP2))
837 if (!isEquivalentOperation(F1I, F2I))
840 assert(F1I->getNumOperands() == F2I->getNumOperands());
841 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) {
842 Value *OpF1 = F1I->getOperand(i);
843 Value *OpF2 = F2I->getOperand(i);
845 if (!enumerate(OpF1, OpF2))
848 if (OpF1->getValueID() != OpF2->getValueID() ||
849 !isEquivalentType(OpF1->getType(), OpF2->getType()))
855 } while (F1I != F1E && F2I != F2E);
857 return F1I == F1E && F2I == F2E;
860 // Test whether the two functions have equivalent behaviour.
861 bool FunctionComparator::compare() {
862 // We need to recheck everything, but check the things that weren't included
863 // in the hash first.
868 if (F1->getAttributes() != F2->getAttributes())
871 if (F1->hasGC() != F2->hasGC())
874 if (F1->hasGC() && F1->getGC() != F2->getGC())
877 if (F1->hasSection() != F2->hasSection())
880 if (F1->hasSection() && F1->getSection() != F2->getSection())
883 if (F1->isVarArg() != F2->isVarArg())
886 // TODO: if it's internal and only used in direct calls, we could handle this
888 if (F1->getCallingConv() != F2->getCallingConv())
891 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType()))
894 assert(F1->arg_size() == F2->arg_size() &&
895 "Identically typed functions have different numbers of args!");
897 // Visit the arguments so that they get enumerated in the order they're
899 for (Function::const_arg_iterator f1i = F1->arg_begin(),
900 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) {
901 if (!enumerate(f1i, f2i))
902 llvm_unreachable("Arguments repeat!");
905 // We do a CFG-ordered walk since the actual ordering of the blocks in the
906 // linked list is immaterial. Our walk starts at the entry block for both
907 // functions, then takes each block from each terminator in order. As an
908 // artifact, this also means that unreachable blocks are ignored.
909 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs;
910 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
912 F1BBs.push_back(&F1->getEntryBlock());
913 F2BBs.push_back(&F2->getEntryBlock());
915 VisitedBBs.insert(F1BBs[0]);
916 while (!F1BBs.empty()) {
917 const BasicBlock *F1BB = F1BBs.pop_back_val();
918 const BasicBlock *F2BB = F2BBs.pop_back_val();
920 if (!enumerate(F1BB, F2BB) || !compare(F1BB, F2BB))
923 const TerminatorInst *F1TI = F1BB->getTerminator();
924 const TerminatorInst *F2TI = F2BB->getTerminator();
926 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors());
927 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) {
928 if (!VisitedBBs.insert(F1TI->getSuccessor(i)))
931 F1BBs.push_back(F1TI->getSuccessor(i));
932 F2BBs.push_back(F2TI->getSuccessor(i));
940 /// MergeFunctions finds functions which will generate identical machine code,
941 /// by considering all pointer types to be equivalent. Once identified,
942 /// MergeFunctions will fold them by replacing a call to one to a call to a
943 /// bitcast of the other.
945 class MergeFunctions : public ModulePass {
949 : ModulePass(ID), HasGlobalAliases(false) {
950 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
953 bool runOnModule(Module &M) override;
956 typedef DenseSet<ComparableFunction> FnSetType;
958 /// A work queue of functions that may have been modified and should be
960 std::vector<WeakVH> Deferred;
962 /// Insert a ComparableFunction into the FnSet, or merge it away if it's
963 /// equal to one that's already present.
964 bool insert(ComparableFunction &NewF);
966 /// Remove a Function from the FnSet and queue it up for a second sweep of
968 void remove(Function *F);
970 /// Find the functions that use this Value and remove them from FnSet and
971 /// queue the functions.
972 void removeUsers(Value *V);
974 /// Replace all direct calls of Old with calls of New. Will bitcast New if
975 /// necessary to make types match.
976 void replaceDirectCallers(Function *Old, Function *New);
978 /// Merge two equivalent functions. Upon completion, G may be deleted, or may
979 /// be converted into a thunk. In either case, it should never be visited
981 void mergeTwoFunctions(Function *F, Function *G);
983 /// Replace G with a thunk or an alias to F. Deletes G.
984 void writeThunkOrAlias(Function *F, Function *G);
986 /// Replace G with a simple tail call to bitcast(F). Also replace direct uses
987 /// of G with bitcast(F). Deletes G.
988 void writeThunk(Function *F, Function *G);
990 /// Replace G with an alias to F. Deletes G.
991 void writeAlias(Function *F, Function *G);
993 /// The set of all distinct functions. Use the insert() and remove() methods
997 /// DataLayout for more accurate GEP comparisons. May be NULL.
998 const DataLayout *DL;
1000 /// Whether or not the target supports global aliases.
1001 bool HasGlobalAliases;
1004 } // end anonymous namespace
1006 char MergeFunctions::ID = 0;
1007 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
1009 ModulePass *llvm::createMergeFunctionsPass() {
1010 return new MergeFunctions();
1013 bool MergeFunctions::runOnModule(Module &M) {
1014 bool Changed = false;
1015 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1016 DL = DLP ? &DLP->getDataLayout() : nullptr;
1018 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
1019 if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage())
1020 Deferred.push_back(WeakVH(I));
1022 FnSet.resize(Deferred.size());
1025 std::vector<WeakVH> Worklist;
1026 Deferred.swap(Worklist);
1028 DEBUG(dbgs() << "size of module: " << M.size() << '\n');
1029 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
1031 // Insert only strong functions and merge them. Strong function merging
1032 // always deletes one of them.
1033 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
1034 E = Worklist.end(); I != E; ++I) {
1036 Function *F = cast<Function>(*I);
1037 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
1038 !F->mayBeOverridden()) {
1039 ComparableFunction CF = ComparableFunction(F, DL);
1040 Changed |= insert(CF);
1044 // Insert only weak functions and merge them. By doing these second we
1045 // create thunks to the strong function when possible. When two weak
1046 // functions are identical, we create a new strong function with two weak
1047 // weak thunks to it which are identical but not mergable.
1048 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
1049 E = Worklist.end(); I != E; ++I) {
1051 Function *F = cast<Function>(*I);
1052 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
1053 F->mayBeOverridden()) {
1054 ComparableFunction CF = ComparableFunction(F, DL);
1055 Changed |= insert(CF);
1058 DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n');
1059 } while (!Deferred.empty());
1066 bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS,
1067 const ComparableFunction &RHS) {
1068 if (LHS.getFunc() == RHS.getFunc() &&
1069 LHS.getHash() == RHS.getHash())
1071 if (!LHS.getFunc() || !RHS.getFunc())
1074 // One of these is a special "underlying pointer comparison only" object.
1075 if (LHS.getDataLayout() == ComparableFunction::LookupOnly ||
1076 RHS.getDataLayout() == ComparableFunction::LookupOnly)
1079 assert(LHS.getDataLayout() == RHS.getDataLayout() &&
1080 "Comparing functions for different targets");
1082 return FunctionComparator(LHS.getDataLayout(), LHS.getFunc(),
1083 RHS.getFunc()).compare();
1086 // Replace direct callers of Old with New.
1087 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
1088 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
1089 for (auto UI = Old->use_begin(), UE = Old->use_end(); UI != UE;) {
1092 CallSite CS(U->getUser());
1093 if (CS && CS.isCallee(U)) {
1094 remove(CS.getInstruction()->getParent()->getParent());
1100 // Replace G with an alias to F if possible, or else a thunk to F. Deletes G.
1101 void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) {
1102 if (HasGlobalAliases && G->hasUnnamedAddr()) {
1103 if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
1104 G->hasWeakLinkage()) {
1113 // Helper for writeThunk,
1114 // Selects proper bitcast operation,
1115 // but a bit simpler then CastInst::getCastOpcode.
1116 static Value *createCast(IRBuilder<false> &Builder, Value *V, Type *DestTy) {
1117 Type *SrcTy = V->getType();
1118 if (SrcTy->isStructTy()) {
1119 assert(DestTy->isStructTy());
1120 assert(SrcTy->getStructNumElements() == DestTy->getStructNumElements());
1121 Value *Result = UndefValue::get(DestTy);
1122 for (unsigned int I = 0, E = SrcTy->getStructNumElements(); I < E; ++I) {
1123 Value *Element = createCast(
1124 Builder, Builder.CreateExtractValue(V, ArrayRef<unsigned int>(I)),
1125 DestTy->getStructElementType(I));
1128 Builder.CreateInsertValue(Result, Element, ArrayRef<unsigned int>(I));
1132 assert(!DestTy->isStructTy());
1133 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
1134 return Builder.CreateIntToPtr(V, DestTy);
1135 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
1136 return Builder.CreatePtrToInt(V, DestTy);
1138 return Builder.CreateBitCast(V, DestTy);
1141 // Replace G with a simple tail call to bitcast(F). Also replace direct uses
1142 // of G with bitcast(F). Deletes G.
1143 void MergeFunctions::writeThunk(Function *F, Function *G) {
1144 if (!G->mayBeOverridden()) {
1145 // Redirect direct callers of G to F.
1146 replaceDirectCallers(G, F);
1149 // If G was internal then we may have replaced all uses of G with F. If so,
1150 // stop here and delete G. There's no need for a thunk.
1151 if (G->hasLocalLinkage() && G->use_empty()) {
1152 G->eraseFromParent();
1156 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
1158 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
1159 IRBuilder<false> Builder(BB);
1161 SmallVector<Value *, 16> Args;
1163 FunctionType *FFTy = F->getFunctionType();
1164 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
1166 Args.push_back(createCast(Builder, (Value*)AI, FFTy->getParamType(i)));
1170 CallInst *CI = Builder.CreateCall(F, Args);
1172 CI->setCallingConv(F->getCallingConv());
1173 if (NewG->getReturnType()->isVoidTy()) {
1174 Builder.CreateRetVoid();
1176 Builder.CreateRet(createCast(Builder, CI, NewG->getReturnType()));
1179 NewG->copyAttributesFrom(G);
1182 G->replaceAllUsesWith(NewG);
1183 G->eraseFromParent();
1185 DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n');
1189 // Replace G with an alias to F and delete G.
1190 void MergeFunctions::writeAlias(Function *F, Function *G) {
1191 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
1192 GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "",
1193 BitcastF, G->getParent());
1194 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
1196 GA->setVisibility(G->getVisibility());
1198 G->replaceAllUsesWith(GA);
1199 G->eraseFromParent();
1201 DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n');
1202 ++NumAliasesWritten;
1205 // Merge two equivalent functions. Upon completion, Function G is deleted.
1206 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) {
1207 if (F->mayBeOverridden()) {
1208 assert(G->mayBeOverridden());
1210 if (HasGlobalAliases) {
1211 // Make them both thunks to the same internal function.
1212 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
1214 H->copyAttributesFrom(F);
1217 F->replaceAllUsesWith(H);
1219 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
1224 F->setAlignment(MaxAlignment);
1225 F->setLinkage(GlobalValue::PrivateLinkage);
1227 // We can't merge them. Instead, pick one and update all direct callers
1228 // to call it and hope that we improve the instruction cache hit rate.
1229 replaceDirectCallers(G, F);
1234 writeThunkOrAlias(F, G);
1237 ++NumFunctionsMerged;
1240 // Insert a ComparableFunction into the FnSet, or merge it away if equal to one
1241 // that was already inserted.
1242 bool MergeFunctions::insert(ComparableFunction &NewF) {
1243 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF);
1244 if (Result.second) {
1245 DEBUG(dbgs() << "Inserting as unique: " << NewF.getFunc()->getName() << '\n');
1249 const ComparableFunction &OldF = *Result.first;
1251 // Don't merge tiny functions, since it can just end up making the function
1253 // FIXME: Should still merge them if they are unnamed_addr and produce an
1255 if (NewF.getFunc()->size() == 1) {
1256 if (NewF.getFunc()->front().size() <= 2) {
1257 DEBUG(dbgs() << NewF.getFunc()->getName()
1258 << " is to small to bother merging\n");
1263 // Never thunk a strong function to a weak function.
1264 assert(!OldF.getFunc()->mayBeOverridden() ||
1265 NewF.getFunc()->mayBeOverridden());
1267 DEBUG(dbgs() << " " << OldF.getFunc()->getName() << " == "
1268 << NewF.getFunc()->getName() << '\n');
1270 Function *DeleteF = NewF.getFunc();
1272 mergeTwoFunctions(OldF.getFunc(), DeleteF);
1276 // Remove a function from FnSet. If it was already in FnSet, add it to Deferred
1277 // so that we'll look at it in the next round.
1278 void MergeFunctions::remove(Function *F) {
1279 // We need to make sure we remove F, not a function "equal" to F per the
1280 // function equality comparator.
1282 // The special "lookup only" ComparableFunction bypasses the expensive
1283 // function comparison in favour of a pointer comparison on the underlying
1285 ComparableFunction CF = ComparableFunction(F, ComparableFunction::LookupOnly);
1286 if (FnSet.erase(CF)) {
1287 DEBUG(dbgs() << "Removed " << F->getName() << " from set and deferred it.\n");
1288 Deferred.push_back(F);
1292 // For each instruction used by the value, remove() the function that contains
1293 // the instruction. This should happen right before a call to RAUW.
1294 void MergeFunctions::removeUsers(Value *V) {
1295 std::vector<Value *> Worklist;
1296 Worklist.push_back(V);
1297 while (!Worklist.empty()) {
1298 Value *V = Worklist.back();
1299 Worklist.pop_back();
1301 for (User *U : V->users()) {
1302 if (Instruction *I = dyn_cast<Instruction>(U)) {
1303 remove(I->getParent()->getParent());
1304 } else if (isa<GlobalValue>(U)) {
1306 } else if (Constant *C = dyn_cast<Constant>(U)) {
1307 for (User *UU : C->users())
1308 Worklist.push_back(UU);