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 // Order relation is defined on set of functions. It was made through
13 // special function comparison procedure that returns
14 // 0 when functions are equal,
15 // -1 when Left function is less than right function, and
16 // 1 for opposite case. We need total-ordering, so we need to maintain
17 // four properties on the functions set:
18 // a <= a (reflexivity)
19 // if a <= b and b <= a then a = b (antisymmetry)
20 // if a <= b and b <= c then a <= c (transitivity).
21 // for all a and b: a <= b or b <= a (totality).
23 // Comparison iterates through each instruction in each basic block.
24 // Functions are kept on binary tree. For each new function F we perform
25 // lookup in binary tree.
26 // In practice it works the following way:
27 // -- We define Function* container class with custom "operator<" (FunctionPtr).
28 // -- "FunctionPtr" instances are stored in std::set collection, so every
29 // std::set::insert operation will give you result in log(N) time.
31 // As an optimization, a hash of the function structure is calculated first, and
32 // two functions are only compared if they have the same hash. This hash is
33 // cheap to compute, and has the property that if function F == G according to
34 // the comparison function, then hash(F) == hash(G). This consistency property
35 // is critical to ensuring all possible merging opportunities are exploited.
36 // Collisions in the hash affect the speed of the pass but not the correctness
37 // or determinism of the resulting transformation.
39 // When a match is found the functions are folded. If both functions are
40 // overridable, we move the functionality into a new internal function and
41 // leave two overridable thunks to it.
43 //===----------------------------------------------------------------------===//
47 // * virtual functions.
49 // Many functions have their address taken by the virtual function table for
50 // the object they belong to. However, as long as it's only used for a lookup
51 // and call, this is irrelevant, and we'd like to fold such functions.
53 // * be smarter about bitcasts.
55 // In order to fold functions, we will sometimes add either bitcast instructions
56 // or bitcast constant expressions. Unfortunately, this can confound further
57 // analysis since the two functions differ where one has a bitcast and the
58 // other doesn't. We should learn to look through bitcasts.
60 // * Compare complex types with pointer types inside.
61 // * Compare cross-reference cases.
62 // * Compare complex expressions.
64 // All the three issues above could be described as ability to prove that
65 // fA == fB == fC == fE == fF == fG in example below:
84 // Simplest cross-reference case (fA <--> fB) was implemented in previous
85 // versions of MergeFunctions, though it presented only in two function pairs
86 // in test-suite (that counts >50k functions)
87 // Though possibility to detect complex cross-referencing (e.g.: A->B->C->D->A)
88 // could cover much more cases.
90 //===----------------------------------------------------------------------===//
92 #include "llvm/Transforms/IPO.h"
93 #include "llvm/ADT/DenseSet.h"
94 #include "llvm/ADT/FoldingSet.h"
95 #include "llvm/ADT/STLExtras.h"
96 #include "llvm/ADT/SmallSet.h"
97 #include "llvm/ADT/Statistic.h"
98 #include "llvm/ADT/Hashing.h"
99 #include "llvm/IR/CallSite.h"
100 #include "llvm/IR/Constants.h"
101 #include "llvm/IR/DataLayout.h"
102 #include "llvm/IR/IRBuilder.h"
103 #include "llvm/IR/InlineAsm.h"
104 #include "llvm/IR/Instructions.h"
105 #include "llvm/IR/LLVMContext.h"
106 #include "llvm/IR/Module.h"
107 #include "llvm/IR/Operator.h"
108 #include "llvm/IR/ValueHandle.h"
109 #include "llvm/IR/ValueMap.h"
110 #include "llvm/Pass.h"
111 #include "llvm/Support/CommandLine.h"
112 #include "llvm/Support/Debug.h"
113 #include "llvm/Support/ErrorHandling.h"
114 #include "llvm/Support/raw_ostream.h"
116 using namespace llvm;
118 #define DEBUG_TYPE "mergefunc"
120 STATISTIC(NumFunctionsMerged, "Number of functions merged");
121 STATISTIC(NumThunksWritten, "Number of thunks generated");
122 STATISTIC(NumAliasesWritten, "Number of aliases generated");
123 STATISTIC(NumDoubleWeak, "Number of new functions created");
125 static cl::opt<unsigned> NumFunctionsForSanityCheck(
127 cl::desc("How many functions in module could be used for "
128 "MergeFunctions pass sanity check. "
129 "'0' disables this check. Works only with '-debug' key."),
130 cl::init(0), cl::Hidden);
134 /// GlobalNumberState assigns an integer to each global value in the program,
135 /// which is used by the comparison routine to order references to globals. This
136 /// state must be preserved throughout the pass, because Functions and other
137 /// globals need to maintain their relative order. Globals are assigned a number
138 /// when they are first visited. This order is deterministic, and so the
139 /// assigned numbers are as well. When two functions are merged, neither number
140 /// is updated. If the symbols are weak, this would be incorrect. If they are
141 /// strong, then one will be replaced at all references to the other, and so
142 /// direct callsites will now see one or the other symbol, and no update is
143 /// necessary. Note that if we were guaranteed unique names, we could just
144 /// compare those, but this would not work for stripped bitcodes or for those
145 /// few symbols without a name.
146 class GlobalNumberState {
147 struct Config : ValueMapConfig<GlobalValue*> {
148 enum { FollowRAUW = false };
150 // Each GlobalValue is mapped to an identifier. The Config ensures when RAUW
151 // occurs, the mapping does not change. Tracking changes is unnecessary, and
152 // also problematic for weak symbols (which may be overwritten).
153 typedef ValueMap<GlobalValue *, uint64_t, Config> ValueNumberMap;
154 ValueNumberMap GlobalNumbers;
155 // The next unused serial number to assign to a global.
158 GlobalNumberState() : GlobalNumbers(), NextNumber(0) {}
159 uint64_t getNumber(GlobalValue* Global) {
160 ValueNumberMap::iterator MapIter;
162 std::tie(MapIter, Inserted) = GlobalNumbers.insert({Global, NextNumber});
165 return MapIter->second;
169 /// FunctionComparator - Compares two functions to determine whether or not
170 /// they will generate machine code with the same behaviour. DataLayout is
171 /// used if available. The comparator always fails conservatively (erring on the
172 /// side of claiming that two functions are different).
173 class FunctionComparator {
175 FunctionComparator(const Function *F1, const Function *F2,
176 GlobalNumberState* GN)
177 : FnL(F1), FnR(F2), GlobalNumbers(GN) {}
179 /// Test whether the two functions have equivalent behaviour.
181 /// Hash a function. Equivalent functions will have the same hash, and unequal
182 /// functions will have different hashes with high probability.
183 typedef uint64_t FunctionHash;
184 static FunctionHash functionHash(Function &);
187 /// Test whether two basic blocks have equivalent behaviour.
188 int cmpBasicBlocks(const BasicBlock *BBL, const BasicBlock *BBR);
190 /// Constants comparison.
191 /// Its analog to lexicographical comparison between hypothetical numbers
193 /// <bitcastability-trait><raw-bit-contents>
195 /// 1. Bitcastability.
196 /// Check whether L's type could be losslessly bitcasted to R's type.
197 /// On this stage method, in case when lossless bitcast is not possible
198 /// method returns -1 or 1, thus also defining which type is greater in
199 /// context of bitcastability.
200 /// Stage 0: If types are equal in terms of cmpTypes, then we can go straight
201 /// to the contents comparison.
202 /// If types differ, remember types comparison result and check
203 /// whether we still can bitcast types.
204 /// Stage 1: Types that satisfies isFirstClassType conditions are always
205 /// greater then others.
206 /// Stage 2: Vector is greater then non-vector.
207 /// If both types are vectors, then vector with greater bitwidth is
209 /// If both types are vectors with the same bitwidth, then types
210 /// are bitcastable, and we can skip other stages, and go to contents
212 /// Stage 3: Pointer types are greater than non-pointers. If both types are
213 /// pointers of the same address space - go to contents comparison.
214 /// Different address spaces: pointer with greater address space is
216 /// Stage 4: Types are neither vectors, nor pointers. And they differ.
217 /// We don't know how to bitcast them. So, we better don't do it,
218 /// and return types comparison result (so it determines the
219 /// relationship among constants we don't know how to bitcast).
221 /// Just for clearance, let's see how the set of constants could look
222 /// on single dimension axis:
224 /// [NFCT], [FCT, "others"], [FCT, pointers], [FCT, vectors]
225 /// Where: NFCT - Not a FirstClassType
226 /// FCT - FirstClassTyp:
228 /// 2. Compare raw contents.
229 /// It ignores types on this stage and only compares bits from L and R.
230 /// Returns 0, if L and R has equivalent contents.
231 /// -1 or 1 if values are different.
233 /// 2.1. If contents are numbers, compare numbers.
234 /// Ints with greater bitwidth are greater. Ints with same bitwidths
235 /// compared by their contents.
236 /// 2.2. "And so on". Just to avoid discrepancies with comments
237 /// perhaps it would be better to read the implementation itself.
238 /// 3. And again about overall picture. Let's look back at how the ordered set
239 /// of constants will look like:
240 /// [NFCT], [FCT, "others"], [FCT, pointers], [FCT, vectors]
242 /// Now look, what could be inside [FCT, "others"], for example:
243 /// [FCT, "others"] =
245 /// [double 0.1], [double 1.23],
246 /// [i32 1], [i32 2],
247 /// { double 1.0 }, ; StructTyID, NumElements = 1
248 /// { i32 1 }, ; StructTyID, NumElements = 1
249 /// { double 1, i32 1 }, ; StructTyID, NumElements = 2
250 /// { i32 1, double 1 } ; StructTyID, NumElements = 2
253 /// Let's explain the order. Float numbers will be less than integers, just
254 /// because of cmpType terms: FloatTyID < IntegerTyID.
255 /// Floats (with same fltSemantics) are sorted according to their value.
256 /// Then you can see integers, and they are, like a floats,
257 /// could be easy sorted among each others.
258 /// The structures. Structures are grouped at the tail, again because of their
259 /// TypeID: StructTyID > IntegerTyID > FloatTyID.
260 /// Structures with greater number of elements are greater. Structures with
261 /// greater elements going first are greater.
262 /// The same logic with vectors, arrays and other possible complex types.
264 /// Bitcastable constants.
265 /// Let's assume, that some constant, belongs to some group of
266 /// "so-called-equal" values with different types, and at the same time
267 /// belongs to another group of constants with equal types
268 /// and "really" equal values.
270 /// Now, prove that this is impossible:
272 /// If constant A with type TyA is bitcastable to B with type TyB, then:
273 /// 1. All constants with equal types to TyA, are bitcastable to B. Since
274 /// those should be vectors (if TyA is vector), pointers
275 /// (if TyA is pointer), or else (if TyA equal to TyB), those types should
277 /// 2. All constants with non-equal, but bitcastable types to TyA, are
278 /// bitcastable to B.
279 /// Once again, just because we allow it to vectors and pointers only.
280 /// This statement could be expanded as below:
281 /// 2.1. All vectors with equal bitwidth to vector A, has equal bitwidth to
282 /// vector B, and thus bitcastable to B as well.
283 /// 2.2. All pointers of the same address space, no matter what they point to,
284 /// bitcastable. So if C is pointer, it could be bitcasted to A and to B.
285 /// So any constant equal or bitcastable to A is equal or bitcastable to B.
288 /// In another words, for pointers and vectors, we ignore top-level type and
289 /// look at their particular properties (bit-width for vectors, and
290 /// address space for pointers).
291 /// If these properties are equal - compare their contents.
292 int cmpConstants(const Constant *L, const Constant *R);
294 /// Compares two global values by number. Uses the GlobalNumbersState to
295 /// identify the same gobals across function calls.
296 int cmpGlobalValues(GlobalValue *L, GlobalValue *R);
298 /// Assign or look up previously assigned numbers for the two values, and
299 /// return whether the numbers are equal. Numbers are assigned in the order
301 /// Comparison order:
302 /// Stage 0: Value that is function itself is always greater then others.
303 /// If left and right values are references to their functions, then
305 /// Stage 1: Constants are greater than non-constants.
306 /// If both left and right are constants, then the result of
307 /// cmpConstants is used as cmpValues result.
308 /// Stage 2: InlineAsm instances are greater than others. If both left and
309 /// right are InlineAsm instances, InlineAsm* pointers casted to
310 /// integers and compared as numbers.
311 /// Stage 3: For all other cases we compare order we meet these values in
312 /// their functions. If right value was met first during scanning,
313 /// then left value is greater.
314 /// In another words, we compare serial numbers, for more details
315 /// see comments for sn_mapL and sn_mapR.
316 int cmpValues(const Value *L, const Value *R);
318 /// Compare two Instructions for equivalence, similar to
319 /// Instruction::isSameOperationAs but with modifications to the type
321 /// Stages are listed in "most significant stage first" order:
322 /// On each stage below, we do comparison between some left and right
323 /// operation parts. If parts are non-equal, we assign parts comparison
324 /// result to the operation comparison result and exit from method.
325 /// Otherwise we proceed to the next stage.
327 /// 1. Operations opcodes. Compared as numbers.
328 /// 2. Number of operands.
329 /// 3. Operation types. Compared with cmpType method.
330 /// 4. Compare operation subclass optional data as stream of bytes:
331 /// just convert it to integers and call cmpNumbers.
332 /// 5. Compare in operation operand types with cmpType in
333 /// most significant operand first order.
334 /// 6. Last stage. Check operations for some specific attributes.
335 /// For example, for Load it would be:
336 /// 6.1.Load: volatile (as boolean flag)
337 /// 6.2.Load: alignment (as integer numbers)
338 /// 6.3.Load: synch-scope (as integer numbers)
339 /// 6.4.Load: range metadata (as integer numbers)
340 /// On this stage its better to see the code, since its not more than 10-15
341 /// strings for particular instruction, and could change sometimes.
342 int cmpOperations(const Instruction *L, const Instruction *R) const;
344 /// Compare two GEPs for equivalent pointer arithmetic.
345 /// Parts to be compared for each comparison stage,
346 /// most significant stage first:
347 /// 1. Address space. As numbers.
348 /// 2. Constant offset, (using GEPOperator::accumulateConstantOffset method).
349 /// 3. Pointer operand type (using cmpType method).
350 /// 4. Number of operands.
351 /// 5. Compare operands, using cmpValues method.
352 int cmpGEPs(const GEPOperator *GEPL, const GEPOperator *GEPR);
353 int cmpGEPs(const GetElementPtrInst *GEPL, const GetElementPtrInst *GEPR) {
354 return cmpGEPs(cast<GEPOperator>(GEPL), cast<GEPOperator>(GEPR));
357 /// cmpType - compares two types,
358 /// defines total ordering among the types set.
361 /// 0 if types are equal,
362 /// -1 if Left is less than Right,
363 /// +1 if Left is greater than Right.
366 /// Comparison is broken onto stages. Like in lexicographical comparison
367 /// stage coming first has higher priority.
368 /// On each explanation stage keep in mind total ordering properties.
370 /// 0. Before comparison we coerce pointer types of 0 address space to
372 /// We also don't bother with same type at left and right, so
373 /// just return 0 in this case.
375 /// 1. If types are of different kind (different type IDs).
376 /// Return result of type IDs comparison, treating them as numbers.
377 /// 2. If types are integers, check that they have the same width. If they
378 /// are vectors, check that they have the same count and subtype.
379 /// 3. Types have the same ID, so check whether they are one of:
388 /// We can treat these types as equal whenever their IDs are same.
389 /// 4. If Left and Right are pointers, return result of address space
390 /// comparison (numbers comparison). We can treat pointer types of same
391 /// address space as equal.
392 /// 5. If types are complex.
393 /// Then both Left and Right are to be expanded and their element types will
394 /// be checked with the same way. If we get Res != 0 on some stage, return it.
395 /// Otherwise return 0.
396 /// 6. For all other cases put llvm_unreachable.
397 int cmpTypes(Type *TyL, Type *TyR) const;
399 int cmpNumbers(uint64_t L, uint64_t R) const;
401 int cmpAPInts(const APInt &L, const APInt &R) const;
402 int cmpAPFloats(const APFloat &L, const APFloat &R) const;
403 int cmpInlineAsm(const InlineAsm *L, const InlineAsm *R) const;
404 int cmpMem(StringRef L, StringRef R) const;
405 int cmpAttrs(const AttributeSet L, const AttributeSet R) const;
407 // The two functions undergoing comparison.
408 const Function *FnL, *FnR;
410 /// Assign serial numbers to values from left function, and values from
413 /// Being comparing functions we need to compare values we meet at left and
415 /// Its easy to sort things out for external values. It just should be
416 /// the same value at left and right.
417 /// But for local values (those were introduced inside function body)
418 /// we have to ensure they were introduced at exactly the same place,
419 /// and plays the same role.
420 /// Let's assign serial number to each value when we meet it first time.
421 /// Values that were met at same place will be with same serial numbers.
422 /// In this case it would be good to explain few points about values assigned
423 /// to BBs and other ways of implementation (see below).
425 /// 1. Safety of BB reordering.
426 /// It's safe to change the order of BasicBlocks in function.
427 /// Relationship with other functions and serial numbering will not be
428 /// changed in this case.
429 /// As follows from FunctionComparator::compare(), we do CFG walk: we start
430 /// from the entry, and then take each terminator. So it doesn't matter how in
431 /// fact BBs are ordered in function. And since cmpValues are called during
432 /// this walk, the numbering depends only on how BBs located inside the CFG.
433 /// So the answer is - yes. We will get the same numbering.
435 /// 2. Impossibility to use dominance properties of values.
436 /// If we compare two instruction operands: first is usage of local
437 /// variable AL from function FL, and second is usage of local variable AR
438 /// from FR, we could compare their origins and check whether they are
439 /// defined at the same place.
440 /// But, we are still not able to compare operands of PHI nodes, since those
441 /// could be operands from further BBs we didn't scan yet.
442 /// So it's impossible to use dominance properties in general.
443 DenseMap<const Value*, int> sn_mapL, sn_mapR;
445 // The global state we will use
446 GlobalNumberState* GlobalNumbers;
450 mutable AssertingVH<Function> F;
451 FunctionComparator::FunctionHash Hash;
453 // Note the hash is recalculated potentially multiple times, but it is cheap.
454 FunctionNode(Function *F)
455 : F(F), Hash(FunctionComparator::functionHash(*F)) {}
456 Function *getFunc() const { return F; }
457 FunctionComparator::FunctionHash getHash() const { return Hash; }
459 /// Replace the reference to the function F by the function G, assuming their
460 /// implementations are equal.
461 void replaceBy(Function *G) const {
465 void release() { F = 0; }
469 int FunctionComparator::cmpNumbers(uint64_t L, uint64_t R) const {
470 if (L < R) return -1;
475 int FunctionComparator::cmpAPInts(const APInt &L, const APInt &R) const {
476 if (int Res = cmpNumbers(L.getBitWidth(), R.getBitWidth()))
478 if (L.ugt(R)) return 1;
479 if (R.ugt(L)) return -1;
483 int FunctionComparator::cmpAPFloats(const APFloat &L, const APFloat &R) const {
484 // TODO: This correctly handles all existing fltSemantics, because they all
485 // have different precisions. This isn't very robust, however, if new types
486 // with different exponent ranges are introduced.
487 const fltSemantics &SL = L.getSemantics(), &SR = R.getSemantics();
488 if (int Res = cmpNumbers(APFloat::semanticsPrecision(SL),
489 APFloat::semanticsPrecision(SR)))
491 return cmpAPInts(L.bitcastToAPInt(), R.bitcastToAPInt());
494 int FunctionComparator::cmpMem(StringRef L, StringRef R) const {
495 // Prevent heavy comparison, compare sizes first.
496 if (int Res = cmpNumbers(L.size(), R.size()))
499 // Compare strings lexicographically only when it is necessary: only when
500 // strings are equal in size.
504 int FunctionComparator::cmpAttrs(const AttributeSet L,
505 const AttributeSet R) const {
506 if (int Res = cmpNumbers(L.getNumSlots(), R.getNumSlots()))
509 for (unsigned i = 0, e = L.getNumSlots(); i != e; ++i) {
510 AttributeSet::iterator LI = L.begin(i), LE = L.end(i), RI = R.begin(i),
512 for (; LI != LE && RI != RE; ++LI, ++RI) {
528 /// Constants comparison:
529 /// 1. Check whether type of L constant could be losslessly bitcasted to R
531 /// 2. Compare constant contents.
532 /// For more details see declaration comments.
533 int FunctionComparator::cmpConstants(const Constant *L, const Constant *R) {
535 Type *TyL = L->getType();
536 Type *TyR = R->getType();
538 // Check whether types are bitcastable. This part is just re-factored
539 // Type::canLosslesslyBitCastTo method, but instead of returning true/false,
540 // we also pack into result which type is "less" for us.
541 int TypesRes = cmpTypes(TyL, TyR);
543 // Types are different, but check whether we can bitcast them.
544 if (!TyL->isFirstClassType()) {
545 if (TyR->isFirstClassType())
547 // Neither TyL nor TyR are values of first class type. Return the result
548 // of comparing the types
551 if (!TyR->isFirstClassType()) {
552 if (TyL->isFirstClassType())
557 // Vector -> Vector conversions are always lossless if the two vector types
558 // have the same size, otherwise not.
559 unsigned TyLWidth = 0;
560 unsigned TyRWidth = 0;
562 if (auto *VecTyL = dyn_cast<VectorType>(TyL))
563 TyLWidth = VecTyL->getBitWidth();
564 if (auto *VecTyR = dyn_cast<VectorType>(TyR))
565 TyRWidth = VecTyR->getBitWidth();
567 if (TyLWidth != TyRWidth)
568 return cmpNumbers(TyLWidth, TyRWidth);
570 // Zero bit-width means neither TyL nor TyR are vectors.
572 PointerType *PTyL = dyn_cast<PointerType>(TyL);
573 PointerType *PTyR = dyn_cast<PointerType>(TyR);
575 unsigned AddrSpaceL = PTyL->getAddressSpace();
576 unsigned AddrSpaceR = PTyR->getAddressSpace();
577 if (int Res = cmpNumbers(AddrSpaceL, AddrSpaceR))
585 // TyL and TyR aren't vectors, nor pointers. We don't know how to
591 // OK, types are bitcastable, now check constant contents.
593 if (L->isNullValue() && R->isNullValue())
595 if (L->isNullValue() && !R->isNullValue())
597 if (!L->isNullValue() && R->isNullValue())
600 auto GlobalValueL = const_cast<GlobalValue*>(dyn_cast<GlobalValue>(L));
601 auto GlobalValueR = const_cast<GlobalValue*>(dyn_cast<GlobalValue>(R));
602 if (GlobalValueL && GlobalValueR) {
603 return cmpGlobalValues(GlobalValueL, GlobalValueR);
606 if (int Res = cmpNumbers(L->getValueID(), R->getValueID()))
609 if (const auto *SeqL = dyn_cast<ConstantDataSequential>(L)) {
610 const auto *SeqR = dyn_cast<ConstantDataSequential>(R);
611 // This handles ConstantDataArray and ConstantDataVector. Note that we
612 // compare the two raw data arrays, which might differ depending on the host
613 // endianness. This isn't a problem though, because the endiness of a module
614 // will affect the order of the constants, but this order is the same
615 // for a given input module and host platform.
616 return cmpMem(SeqL->getRawDataValues(), SeqR->getRawDataValues());
619 switch (L->getValueID()) {
620 case Value::UndefValueVal: return TypesRes;
621 case Value::ConstantIntVal: {
622 const APInt &LInt = cast<ConstantInt>(L)->getValue();
623 const APInt &RInt = cast<ConstantInt>(R)->getValue();
624 return cmpAPInts(LInt, RInt);
626 case Value::ConstantFPVal: {
627 const APFloat &LAPF = cast<ConstantFP>(L)->getValueAPF();
628 const APFloat &RAPF = cast<ConstantFP>(R)->getValueAPF();
629 return cmpAPFloats(LAPF, RAPF);
631 case Value::ConstantArrayVal: {
632 const ConstantArray *LA = cast<ConstantArray>(L);
633 const ConstantArray *RA = cast<ConstantArray>(R);
634 uint64_t NumElementsL = cast<ArrayType>(TyL)->getNumElements();
635 uint64_t NumElementsR = cast<ArrayType>(TyR)->getNumElements();
636 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
638 for (uint64_t i = 0; i < NumElementsL; ++i) {
639 if (int Res = cmpConstants(cast<Constant>(LA->getOperand(i)),
640 cast<Constant>(RA->getOperand(i))))
645 case Value::ConstantStructVal: {
646 const ConstantStruct *LS = cast<ConstantStruct>(L);
647 const ConstantStruct *RS = cast<ConstantStruct>(R);
648 unsigned NumElementsL = cast<StructType>(TyL)->getNumElements();
649 unsigned NumElementsR = cast<StructType>(TyR)->getNumElements();
650 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
652 for (unsigned i = 0; i != NumElementsL; ++i) {
653 if (int Res = cmpConstants(cast<Constant>(LS->getOperand(i)),
654 cast<Constant>(RS->getOperand(i))))
659 case Value::ConstantVectorVal: {
660 const ConstantVector *LV = cast<ConstantVector>(L);
661 const ConstantVector *RV = cast<ConstantVector>(R);
662 unsigned NumElementsL = cast<VectorType>(TyL)->getNumElements();
663 unsigned NumElementsR = cast<VectorType>(TyR)->getNumElements();
664 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
666 for (uint64_t i = 0; i < NumElementsL; ++i) {
667 if (int Res = cmpConstants(cast<Constant>(LV->getOperand(i)),
668 cast<Constant>(RV->getOperand(i))))
673 case Value::ConstantExprVal: {
674 const ConstantExpr *LE = cast<ConstantExpr>(L);
675 const ConstantExpr *RE = cast<ConstantExpr>(R);
676 unsigned NumOperandsL = LE->getNumOperands();
677 unsigned NumOperandsR = RE->getNumOperands();
678 if (int Res = cmpNumbers(NumOperandsL, NumOperandsR))
680 for (unsigned i = 0; i < NumOperandsL; ++i) {
681 if (int Res = cmpConstants(cast<Constant>(LE->getOperand(i)),
682 cast<Constant>(RE->getOperand(i))))
687 case Value::BlockAddressVal: {
688 // FIXME: This still uses a pointer comparison. It isn't clear how to remove
689 // this. This only affects programs which take BlockAddresses and store them
690 // as constants, which is limited to interepreters, etc.
691 return cmpNumbers((uint64_t)L, (uint64_t)R);
693 default: // Unknown constant, abort.
694 DEBUG(dbgs() << "Looking at valueID " << L->getValueID() << "\n");
695 llvm_unreachable("Constant ValueID not recognized.");
700 int FunctionComparator::cmpGlobalValues(GlobalValue *L, GlobalValue* R) {
701 return cmpNumbers(GlobalNumbers->getNumber(L), GlobalNumbers->getNumber(R));
704 /// cmpType - compares two types,
705 /// defines total ordering among the types set.
706 /// See method declaration comments for more details.
707 int FunctionComparator::cmpTypes(Type *TyL, Type *TyR) const {
709 PointerType *PTyL = dyn_cast<PointerType>(TyL);
710 PointerType *PTyR = dyn_cast<PointerType>(TyR);
712 const DataLayout &DL = FnL->getParent()->getDataLayout();
713 if (PTyL && PTyL->getAddressSpace() == 0)
714 TyL = DL.getIntPtrType(TyL);
715 if (PTyR && PTyR->getAddressSpace() == 0)
716 TyR = DL.getIntPtrType(TyR);
721 if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
724 switch (TyL->getTypeID()) {
726 llvm_unreachable("Unknown type!");
727 // Fall through in Release mode.
728 case Type::IntegerTyID:
729 return cmpNumbers(cast<IntegerType>(TyL)->getBitWidth(),
730 cast<IntegerType>(TyR)->getBitWidth());
731 case Type::VectorTyID: {
732 VectorType *VTyL = cast<VectorType>(TyL), *VTyR = cast<VectorType>(TyR);
733 if (int Res = cmpNumbers(VTyL->getNumElements(), VTyR->getNumElements()))
735 return cmpTypes(VTyL->getElementType(), VTyR->getElementType());
737 // TyL == TyR would have returned true earlier, because types are uniqued.
739 case Type::FloatTyID:
740 case Type::DoubleTyID:
741 case Type::X86_FP80TyID:
742 case Type::FP128TyID:
743 case Type::PPC_FP128TyID:
744 case Type::LabelTyID:
745 case Type::MetadataTyID:
746 case Type::TokenTyID:
749 case Type::PointerTyID: {
750 assert(PTyL && PTyR && "Both types must be pointers here.");
751 return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
754 case Type::StructTyID: {
755 StructType *STyL = cast<StructType>(TyL);
756 StructType *STyR = cast<StructType>(TyR);
757 if (STyL->getNumElements() != STyR->getNumElements())
758 return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
760 if (STyL->isPacked() != STyR->isPacked())
761 return cmpNumbers(STyL->isPacked(), STyR->isPacked());
763 for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
764 if (int Res = cmpTypes(STyL->getElementType(i), STyR->getElementType(i)))
770 case Type::FunctionTyID: {
771 FunctionType *FTyL = cast<FunctionType>(TyL);
772 FunctionType *FTyR = cast<FunctionType>(TyR);
773 if (FTyL->getNumParams() != FTyR->getNumParams())
774 return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
776 if (FTyL->isVarArg() != FTyR->isVarArg())
777 return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
779 if (int Res = cmpTypes(FTyL->getReturnType(), FTyR->getReturnType()))
782 for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
783 if (int Res = cmpTypes(FTyL->getParamType(i), FTyR->getParamType(i)))
789 case Type::ArrayTyID: {
790 ArrayType *ATyL = cast<ArrayType>(TyL);
791 ArrayType *ATyR = cast<ArrayType>(TyR);
792 if (ATyL->getNumElements() != ATyR->getNumElements())
793 return cmpNumbers(ATyL->getNumElements(), ATyR->getNumElements());
794 return cmpTypes(ATyL->getElementType(), ATyR->getElementType());
799 // Determine whether the two operations are the same except that pointer-to-A
800 // and pointer-to-B are equivalent. This should be kept in sync with
801 // Instruction::isSameOperationAs.
802 // Read method declaration comments for more details.
803 int FunctionComparator::cmpOperations(const Instruction *L,
804 const Instruction *R) const {
805 // Differences from Instruction::isSameOperationAs:
806 // * replace type comparison with calls to isEquivalentType.
807 // * we test for I->hasSameSubclassOptionalData (nuw/nsw/tail) at the top
808 // * because of the above, we don't test for the tail bit on calls later on
809 if (int Res = cmpNumbers(L->getOpcode(), R->getOpcode()))
812 if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
815 if (int Res = cmpTypes(L->getType(), R->getType()))
818 if (int Res = cmpNumbers(L->getRawSubclassOptionalData(),
819 R->getRawSubclassOptionalData()))
822 if (const AllocaInst *AI = dyn_cast<AllocaInst>(L)) {
823 if (int Res = cmpTypes(AI->getAllocatedType(),
824 cast<AllocaInst>(R)->getAllocatedType()))
827 cmpNumbers(AI->getAlignment(), cast<AllocaInst>(R)->getAlignment()))
831 // We have two instructions of identical opcode and #operands. Check to see
832 // if all operands are the same type
833 for (unsigned i = 0, e = L->getNumOperands(); i != e; ++i) {
835 cmpTypes(L->getOperand(i)->getType(), R->getOperand(i)->getType()))
839 // Check special state that is a part of some instructions.
840 if (const LoadInst *LI = dyn_cast<LoadInst>(L)) {
841 if (int Res = cmpNumbers(LI->isVolatile(), cast<LoadInst>(R)->isVolatile()))
844 cmpNumbers(LI->getAlignment(), cast<LoadInst>(R)->getAlignment()))
847 cmpNumbers(LI->getOrdering(), cast<LoadInst>(R)->getOrdering()))
850 cmpNumbers(LI->getSynchScope(), cast<LoadInst>(R)->getSynchScope()))
852 return cmpNumbers((uint64_t)LI->getMetadata(LLVMContext::MD_range),
853 (uint64_t)cast<LoadInst>(R)->getMetadata(LLVMContext::MD_range));
855 if (const StoreInst *SI = dyn_cast<StoreInst>(L)) {
857 cmpNumbers(SI->isVolatile(), cast<StoreInst>(R)->isVolatile()))
860 cmpNumbers(SI->getAlignment(), cast<StoreInst>(R)->getAlignment()))
863 cmpNumbers(SI->getOrdering(), cast<StoreInst>(R)->getOrdering()))
865 return cmpNumbers(SI->getSynchScope(), cast<StoreInst>(R)->getSynchScope());
867 if (const CmpInst *CI = dyn_cast<CmpInst>(L))
868 return cmpNumbers(CI->getPredicate(), cast<CmpInst>(R)->getPredicate());
869 if (const CallInst *CI = dyn_cast<CallInst>(L)) {
870 if (int Res = cmpNumbers(CI->getCallingConv(),
871 cast<CallInst>(R)->getCallingConv()))
874 cmpAttrs(CI->getAttributes(), cast<CallInst>(R)->getAttributes()))
877 (uint64_t)CI->getMetadata(LLVMContext::MD_range),
878 (uint64_t)cast<CallInst>(R)->getMetadata(LLVMContext::MD_range));
880 if (const InvokeInst *CI = dyn_cast<InvokeInst>(L)) {
881 if (int Res = cmpNumbers(CI->getCallingConv(),
882 cast<InvokeInst>(R)->getCallingConv()))
885 cmpAttrs(CI->getAttributes(), cast<InvokeInst>(R)->getAttributes()))
888 (uint64_t)CI->getMetadata(LLVMContext::MD_range),
889 (uint64_t)cast<InvokeInst>(R)->getMetadata(LLVMContext::MD_range));
891 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(L)) {
892 ArrayRef<unsigned> LIndices = IVI->getIndices();
893 ArrayRef<unsigned> RIndices = cast<InsertValueInst>(R)->getIndices();
894 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
896 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
897 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
901 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(L)) {
902 ArrayRef<unsigned> LIndices = EVI->getIndices();
903 ArrayRef<unsigned> RIndices = cast<ExtractValueInst>(R)->getIndices();
904 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
906 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
907 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
911 if (const FenceInst *FI = dyn_cast<FenceInst>(L)) {
913 cmpNumbers(FI->getOrdering(), cast<FenceInst>(R)->getOrdering()))
915 return cmpNumbers(FI->getSynchScope(), cast<FenceInst>(R)->getSynchScope());
918 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(L)) {
919 if (int Res = cmpNumbers(CXI->isVolatile(),
920 cast<AtomicCmpXchgInst>(R)->isVolatile()))
922 if (int Res = cmpNumbers(CXI->isWeak(),
923 cast<AtomicCmpXchgInst>(R)->isWeak()))
925 if (int Res = cmpNumbers(CXI->getSuccessOrdering(),
926 cast<AtomicCmpXchgInst>(R)->getSuccessOrdering()))
928 if (int Res = cmpNumbers(CXI->getFailureOrdering(),
929 cast<AtomicCmpXchgInst>(R)->getFailureOrdering()))
931 return cmpNumbers(CXI->getSynchScope(),
932 cast<AtomicCmpXchgInst>(R)->getSynchScope());
934 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(L)) {
935 if (int Res = cmpNumbers(RMWI->getOperation(),
936 cast<AtomicRMWInst>(R)->getOperation()))
938 if (int Res = cmpNumbers(RMWI->isVolatile(),
939 cast<AtomicRMWInst>(R)->isVolatile()))
941 if (int Res = cmpNumbers(RMWI->getOrdering(),
942 cast<AtomicRMWInst>(R)->getOrdering()))
944 return cmpNumbers(RMWI->getSynchScope(),
945 cast<AtomicRMWInst>(R)->getSynchScope());
950 // Determine whether two GEP operations perform the same underlying arithmetic.
951 // Read method declaration comments for more details.
952 int FunctionComparator::cmpGEPs(const GEPOperator *GEPL,
953 const GEPOperator *GEPR) {
955 unsigned int ASL = GEPL->getPointerAddressSpace();
956 unsigned int ASR = GEPR->getPointerAddressSpace();
958 if (int Res = cmpNumbers(ASL, ASR))
961 // When we have target data, we can reduce the GEP down to the value in bytes
962 // added to the address.
963 const DataLayout &DL = FnL->getParent()->getDataLayout();
964 unsigned BitWidth = DL.getPointerSizeInBits(ASL);
965 APInt OffsetL(BitWidth, 0), OffsetR(BitWidth, 0);
966 if (GEPL->accumulateConstantOffset(DL, OffsetL) &&
967 GEPR->accumulateConstantOffset(DL, OffsetR))
968 return cmpAPInts(OffsetL, OffsetR);
969 if (int Res = cmpTypes(GEPL->getPointerOperand()->getType(),
970 GEPR->getPointerOperand()->getType()))
973 if (int Res = cmpNumbers(GEPL->getNumOperands(), GEPR->getNumOperands()))
976 for (unsigned i = 0, e = GEPL->getNumOperands(); i != e; ++i) {
977 if (int Res = cmpValues(GEPL->getOperand(i), GEPR->getOperand(i)))
984 int FunctionComparator::cmpInlineAsm(const InlineAsm *L,
985 const InlineAsm *R) const {
986 // InlineAsm's are uniqued. If they are the same pointer, obviously they are
987 // the same, otherwise compare the fields.
990 if (int Res = cmpTypes(L->getFunctionType(), R->getFunctionType()))
992 if (int Res = cmpMem(L->getAsmString(), R->getAsmString()))
994 if (int Res = cmpMem(L->getConstraintString(), R->getConstraintString()))
996 if (int Res = cmpNumbers(L->hasSideEffects(), R->hasSideEffects()))
998 if (int Res = cmpNumbers(L->isAlignStack(), R->isAlignStack()))
1000 if (int Res = cmpNumbers(L->getDialect(), R->getDialect()))
1002 llvm_unreachable("InlineAsm blocks were not uniqued.");
1006 /// Compare two values used by the two functions under pair-wise comparison. If
1007 /// this is the first time the values are seen, they're added to the mapping so
1008 /// that we will detect mismatches on next use.
1009 /// See comments in declaration for more details.
1010 int FunctionComparator::cmpValues(const Value *L, const Value *R) {
1011 // Catch self-reference case.
1023 const Constant *ConstL = dyn_cast<Constant>(L);
1024 const Constant *ConstR = dyn_cast<Constant>(R);
1025 if (ConstL && ConstR) {
1028 return cmpConstants(ConstL, ConstR);
1036 const InlineAsm *InlineAsmL = dyn_cast<InlineAsm>(L);
1037 const InlineAsm *InlineAsmR = dyn_cast<InlineAsm>(R);
1039 if (InlineAsmL && InlineAsmR)
1040 return cmpInlineAsm(InlineAsmL, InlineAsmR);
1046 auto LeftSN = sn_mapL.insert(std::make_pair(L, sn_mapL.size())),
1047 RightSN = sn_mapR.insert(std::make_pair(R, sn_mapR.size()));
1049 return cmpNumbers(LeftSN.first->second, RightSN.first->second);
1051 // Test whether two basic blocks have equivalent behaviour.
1052 int FunctionComparator::cmpBasicBlocks(const BasicBlock *BBL,
1053 const BasicBlock *BBR) {
1054 BasicBlock::const_iterator InstL = BBL->begin(), InstLE = BBL->end();
1055 BasicBlock::const_iterator InstR = BBR->begin(), InstRE = BBR->end();
1058 if (int Res = cmpValues(InstL, InstR))
1061 const GetElementPtrInst *GEPL = dyn_cast<GetElementPtrInst>(InstL);
1062 const GetElementPtrInst *GEPR = dyn_cast<GetElementPtrInst>(InstR);
1071 cmpValues(GEPL->getPointerOperand(), GEPR->getPointerOperand()))
1073 if (int Res = cmpGEPs(GEPL, GEPR))
1076 if (int Res = cmpOperations(InstL, InstR))
1078 assert(InstL->getNumOperands() == InstR->getNumOperands());
1080 for (unsigned i = 0, e = InstL->getNumOperands(); i != e; ++i) {
1081 Value *OpL = InstL->getOperand(i);
1082 Value *OpR = InstR->getOperand(i);
1083 if (int Res = cmpValues(OpL, OpR))
1085 // cmpValues should ensure this is true.
1086 assert(cmpTypes(OpL->getType(), OpR->getType()) == 0);
1091 } while (InstL != InstLE && InstR != InstRE);
1093 if (InstL != InstLE && InstR == InstRE)
1095 if (InstL == InstLE && InstR != InstRE)
1100 // Test whether the two functions have equivalent behaviour.
1101 int FunctionComparator::compare() {
1106 if (int Res = cmpAttrs(FnL->getAttributes(), FnR->getAttributes()))
1109 if (int Res = cmpNumbers(FnL->hasGC(), FnR->hasGC()))
1113 if (int Res = cmpMem(FnL->getGC(), FnR->getGC()))
1117 if (int Res = cmpNumbers(FnL->hasSection(), FnR->hasSection()))
1120 if (FnL->hasSection()) {
1121 if (int Res = cmpMem(FnL->getSection(), FnR->getSection()))
1125 if (int Res = cmpNumbers(FnL->isVarArg(), FnR->isVarArg()))
1128 // TODO: if it's internal and only used in direct calls, we could handle this
1130 if (int Res = cmpNumbers(FnL->getCallingConv(), FnR->getCallingConv()))
1133 if (int Res = cmpTypes(FnL->getFunctionType(), FnR->getFunctionType()))
1136 assert(FnL->arg_size() == FnR->arg_size() &&
1137 "Identically typed functions have different numbers of args!");
1139 // Visit the arguments so that they get enumerated in the order they're
1141 for (Function::const_arg_iterator ArgLI = FnL->arg_begin(),
1142 ArgRI = FnR->arg_begin(),
1143 ArgLE = FnL->arg_end();
1144 ArgLI != ArgLE; ++ArgLI, ++ArgRI) {
1145 if (cmpValues(ArgLI, ArgRI) != 0)
1146 llvm_unreachable("Arguments repeat!");
1149 // We do a CFG-ordered walk since the actual ordering of the blocks in the
1150 // linked list is immaterial. Our walk starts at the entry block for both
1151 // functions, then takes each block from each terminator in order. As an
1152 // artifact, this also means that unreachable blocks are ignored.
1153 SmallVector<const BasicBlock *, 8> FnLBBs, FnRBBs;
1154 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
1156 FnLBBs.push_back(&FnL->getEntryBlock());
1157 FnRBBs.push_back(&FnR->getEntryBlock());
1159 VisitedBBs.insert(FnLBBs[0]);
1160 while (!FnLBBs.empty()) {
1161 const BasicBlock *BBL = FnLBBs.pop_back_val();
1162 const BasicBlock *BBR = FnRBBs.pop_back_val();
1164 if (int Res = cmpValues(BBL, BBR))
1167 if (int Res = cmpBasicBlocks(BBL, BBR))
1170 const TerminatorInst *TermL = BBL->getTerminator();
1171 const TerminatorInst *TermR = BBR->getTerminator();
1173 assert(TermL->getNumSuccessors() == TermR->getNumSuccessors());
1174 for (unsigned i = 0, e = TermL->getNumSuccessors(); i != e; ++i) {
1175 if (!VisitedBBs.insert(TermL->getSuccessor(i)).second)
1178 FnLBBs.push_back(TermL->getSuccessor(i));
1179 FnRBBs.push_back(TermR->getSuccessor(i));
1185 // Accumulate the hash of a sequence of 64-bit integers. This is similar to a
1186 // hash of a sequence of 64bit ints, but the entire input does not need to be
1187 // available at once. This interface is necessary for functionHash because it
1188 // needs to accumulate the hash as the structure of the function is traversed
1189 // without saving these values to an intermediate buffer. This form of hashing
1190 // is not often needed, as usually the object to hash is just read from a
1192 class HashAccumulator64 {
1195 // Initialize to random constant, so the state isn't zero.
1196 HashAccumulator64() { Hash = 0x6acaa36bef8325c5ULL; }
1197 void add(uint64_t V) {
1198 Hash = llvm::hashing::detail::hash_16_bytes(Hash, V);
1200 // No finishing is required, because the entire hash value is used.
1201 uint64_t getHash() { return Hash; }
1204 // A function hash is calculated by considering only the number of arguments and
1205 // whether a function is varargs, the order of basic blocks (given by the
1206 // successors of each basic block in depth first order), and the order of
1207 // opcodes of each instruction within each of these basic blocks. This mirrors
1208 // the strategy compare() uses to compare functions by walking the BBs in depth
1209 // first order and comparing each instruction in sequence. Because this hash
1210 // does not look at the operands, it is insensitive to things such as the
1211 // target of calls and the constants used in the function, which makes it useful
1212 // when possibly merging functions which are the same modulo constants and call
1214 FunctionComparator::FunctionHash FunctionComparator::functionHash(Function &F) {
1215 HashAccumulator64 H;
1216 H.add(F.isVarArg());
1217 H.add(F.arg_size());
1219 SmallVector<const BasicBlock *, 8> BBs;
1220 SmallSet<const BasicBlock *, 16> VisitedBBs;
1222 // Walk the blocks in the same order as FunctionComparator::cmpBasicBlocks(),
1223 // accumulating the hash of the function "structure." (BB and opcode sequence)
1224 BBs.push_back(&F.getEntryBlock());
1225 VisitedBBs.insert(BBs[0]);
1226 while (!BBs.empty()) {
1227 const BasicBlock *BB = BBs.pop_back_val();
1228 // This random value acts as a block header, as otherwise the partition of
1229 // opcodes into BBs wouldn't affect the hash, only the order of the opcodes
1231 for (auto &Inst : *BB) {
1232 H.add(Inst.getOpcode());
1234 const TerminatorInst *Term = BB->getTerminator();
1235 for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) {
1236 if (!VisitedBBs.insert(Term->getSuccessor(i)).second)
1238 BBs.push_back(Term->getSuccessor(i));
1247 /// MergeFunctions finds functions which will generate identical machine code,
1248 /// by considering all pointer types to be equivalent. Once identified,
1249 /// MergeFunctions will fold them by replacing a call to one to a call to a
1250 /// bitcast of the other.
1252 class MergeFunctions : public ModulePass {
1256 : ModulePass(ID), FnTree(FunctionNodeCmp(&GlobalNumbers)),
1257 HasGlobalAliases(false) {
1258 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
1261 bool runOnModule(Module &M) override;
1264 // The function comparison operator is provided here so that FunctionNodes do
1265 // not need to become larger with another pointer.
1266 class FunctionNodeCmp {
1267 GlobalNumberState* GlobalNumbers;
1269 FunctionNodeCmp(GlobalNumberState* GN) : GlobalNumbers(GN) {}
1270 bool operator()(const FunctionNode &LHS, const FunctionNode &RHS) const {
1271 // Order first by hashes, then full function comparison.
1272 if (LHS.getHash() != RHS.getHash())
1273 return LHS.getHash() < RHS.getHash();
1274 FunctionComparator FCmp(LHS.getFunc(), RHS.getFunc(), GlobalNumbers);
1275 return FCmp.compare() == -1;
1278 typedef std::set<FunctionNode, FunctionNodeCmp> FnTreeType;
1280 GlobalNumberState GlobalNumbers;
1282 /// A work queue of functions that may have been modified and should be
1284 std::vector<WeakVH> Deferred;
1286 /// Checks the rules of order relation introduced among functions set.
1287 /// Returns true, if sanity check has been passed, and false if failed.
1288 bool doSanityCheck(std::vector<WeakVH> &Worklist);
1290 /// Insert a ComparableFunction into the FnTree, or merge it away if it's
1291 /// equal to one that's already present.
1292 bool insert(Function *NewFunction);
1294 /// Remove a Function from the FnTree and queue it up for a second sweep of
1296 void remove(Function *F);
1298 /// Find the functions that use this Value and remove them from FnTree and
1299 /// queue the functions.
1300 void removeUsers(Value *V);
1302 /// Replace all direct calls of Old with calls of New. Will bitcast New if
1303 /// necessary to make types match.
1304 void replaceDirectCallers(Function *Old, Function *New);
1306 /// Merge two equivalent functions. Upon completion, G may be deleted, or may
1307 /// be converted into a thunk. In either case, it should never be visited
1309 void mergeTwoFunctions(Function *F, Function *G);
1311 /// Replace G with a thunk or an alias to F. Deletes G.
1312 void writeThunkOrAlias(Function *F, Function *G);
1314 /// Replace G with a simple tail call to bitcast(F). Also replace direct uses
1315 /// of G with bitcast(F). Deletes G.
1316 void writeThunk(Function *F, Function *G);
1318 /// Replace G with an alias to F. Deletes G.
1319 void writeAlias(Function *F, Function *G);
1321 /// Replace function F with function G in the function tree.
1322 void replaceFunctionInTree(FnTreeType::iterator &IterToF, Function *G);
1324 /// The set of all distinct functions. Use the insert() and remove() methods
1328 /// Whether or not the target supports global aliases.
1329 bool HasGlobalAliases;
1332 } // end anonymous namespace
1334 char MergeFunctions::ID = 0;
1335 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
1337 ModulePass *llvm::createMergeFunctionsPass() {
1338 return new MergeFunctions();
1341 bool MergeFunctions::doSanityCheck(std::vector<WeakVH> &Worklist) {
1342 if (const unsigned Max = NumFunctionsForSanityCheck) {
1343 unsigned TripleNumber = 0;
1346 dbgs() << "MERGEFUNC-SANITY: Started for first " << Max << " functions.\n";
1349 for (std::vector<WeakVH>::iterator I = Worklist.begin(), E = Worklist.end();
1350 I != E && i < Max; ++I, ++i) {
1352 for (std::vector<WeakVH>::iterator J = I; J != E && j < Max; ++J, ++j) {
1353 Function *F1 = cast<Function>(*I);
1354 Function *F2 = cast<Function>(*J);
1355 int Res1 = FunctionComparator(F1, F2, &GlobalNumbers).compare();
1356 int Res2 = FunctionComparator(F2, F1, &GlobalNumbers).compare();
1358 // If F1 <= F2, then F2 >= F1, otherwise report failure.
1359 if (Res1 != -Res2) {
1360 dbgs() << "MERGEFUNC-SANITY: Non-symmetric; triple: " << TripleNumber
1371 for (std::vector<WeakVH>::iterator K = J; K != E && k < Max;
1372 ++k, ++K, ++TripleNumber) {
1376 Function *F3 = cast<Function>(*K);
1377 int Res3 = FunctionComparator(F1, F3, &GlobalNumbers).compare();
1378 int Res4 = FunctionComparator(F2, F3, &GlobalNumbers).compare();
1380 bool Transitive = true;
1382 if (Res1 != 0 && Res1 == Res4) {
1383 // F1 > F2, F2 > F3 => F1 > F3
1384 Transitive = Res3 == Res1;
1385 } else if (Res3 != 0 && Res3 == -Res4) {
1386 // F1 > F3, F3 > F2 => F1 > F2
1387 Transitive = Res3 == Res1;
1388 } else if (Res4 != 0 && -Res3 == Res4) {
1389 // F2 > F3, F3 > F1 => F2 > F1
1390 Transitive = Res4 == -Res1;
1394 dbgs() << "MERGEFUNC-SANITY: Non-transitive; triple: "
1395 << TripleNumber << "\n";
1396 dbgs() << "Res1, Res3, Res4: " << Res1 << ", " << Res3 << ", "
1407 dbgs() << "MERGEFUNC-SANITY: " << (Valid ? "Passed." : "Failed.") << "\n";
1413 bool MergeFunctions::runOnModule(Module &M) {
1414 bool Changed = false;
1416 // All functions in the module, ordered by hash. Functions with a unique
1417 // hash value are easily eliminated.
1418 std::vector<std::pair<FunctionComparator::FunctionHash, Function *>>
1420 for (Function &Func : M) {
1421 if (!Func.isDeclaration() && !Func.hasAvailableExternallyLinkage()) {
1422 HashedFuncs.push_back({FunctionComparator::functionHash(Func), &Func});
1427 HashedFuncs.begin(), HashedFuncs.end(),
1428 [](const std::pair<FunctionComparator::FunctionHash, Function *> &a,
1429 const std::pair<FunctionComparator::FunctionHash, Function *> &b) {
1430 return a.first < b.first;
1433 auto S = HashedFuncs.begin();
1434 for (auto I = HashedFuncs.begin(), IE = HashedFuncs.end(); I != IE; ++I) {
1435 // If the hash value matches the previous value or the next one, we must
1436 // consider merging it. Otherwise it is dropped and never considered again.
1437 if ((I != S && std::prev(I)->first == I->first) ||
1438 (std::next(I) != IE && std::next(I)->first == I->first) ) {
1439 Deferred.push_back(WeakVH(I->second));
1444 std::vector<WeakVH> Worklist;
1445 Deferred.swap(Worklist);
1447 DEBUG(doSanityCheck(Worklist));
1449 DEBUG(dbgs() << "size of module: " << M.size() << '\n');
1450 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
1452 // Insert only strong functions and merge them. Strong function merging
1453 // always deletes one of them.
1454 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
1455 E = Worklist.end(); I != E; ++I) {
1457 Function *F = cast<Function>(*I);
1458 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
1459 !F->mayBeOverridden()) {
1460 Changed |= insert(F);
1464 // Insert only weak functions and merge them. By doing these second we
1465 // create thunks to the strong function when possible. When two weak
1466 // functions are identical, we create a new strong function with two weak
1467 // weak thunks to it which are identical but not mergable.
1468 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
1469 E = Worklist.end(); I != E; ++I) {
1471 Function *F = cast<Function>(*I);
1472 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
1473 F->mayBeOverridden()) {
1474 Changed |= insert(F);
1477 DEBUG(dbgs() << "size of FnTree: " << FnTree.size() << '\n');
1478 } while (!Deferred.empty());
1485 // Replace direct callers of Old with New.
1486 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
1487 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
1488 for (auto UI = Old->use_begin(), UE = Old->use_end(); UI != UE;) {
1491 CallSite CS(U->getUser());
1492 if (CS && CS.isCallee(U)) {
1493 // Transfer the called function's attributes to the call site. Due to the
1494 // bitcast we will 'loose' ABI changing attributes because the 'called
1495 // function' is no longer a Function* but the bitcast. Code that looks up
1496 // the attributes from the called function will fail.
1497 auto &Context = New->getContext();
1498 auto NewFuncAttrs = New->getAttributes();
1499 auto CallSiteAttrs = CS.getAttributes();
1501 CallSiteAttrs = CallSiteAttrs.addAttributes(
1502 Context, AttributeSet::ReturnIndex, NewFuncAttrs.getRetAttributes());
1504 for (unsigned argIdx = 0; argIdx < CS.arg_size(); argIdx++) {
1505 AttributeSet Attrs = NewFuncAttrs.getParamAttributes(argIdx);
1506 if (Attrs.getNumSlots())
1507 CallSiteAttrs = CallSiteAttrs.addAttributes(Context, argIdx, Attrs);
1510 CS.setAttributes(CallSiteAttrs);
1512 remove(CS.getInstruction()->getParent()->getParent());
1518 // Replace G with an alias to F if possible, or else a thunk to F. Deletes G.
1519 void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) {
1520 if (HasGlobalAliases && G->hasUnnamedAddr()) {
1521 if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
1522 G->hasWeakLinkage()) {
1531 // Helper for writeThunk,
1532 // Selects proper bitcast operation,
1533 // but a bit simpler then CastInst::getCastOpcode.
1534 static Value *createCast(IRBuilder<false> &Builder, Value *V, Type *DestTy) {
1535 Type *SrcTy = V->getType();
1536 if (SrcTy->isStructTy()) {
1537 assert(DestTy->isStructTy());
1538 assert(SrcTy->getStructNumElements() == DestTy->getStructNumElements());
1539 Value *Result = UndefValue::get(DestTy);
1540 for (unsigned int I = 0, E = SrcTy->getStructNumElements(); I < E; ++I) {
1541 Value *Element = createCast(
1542 Builder, Builder.CreateExtractValue(V, makeArrayRef(I)),
1543 DestTy->getStructElementType(I));
1546 Builder.CreateInsertValue(Result, Element, makeArrayRef(I));
1550 assert(!DestTy->isStructTy());
1551 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
1552 return Builder.CreateIntToPtr(V, DestTy);
1553 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
1554 return Builder.CreatePtrToInt(V, DestTy);
1556 return Builder.CreateBitCast(V, DestTy);
1559 // Replace G with a simple tail call to bitcast(F). Also replace direct uses
1560 // of G with bitcast(F). Deletes G.
1561 void MergeFunctions::writeThunk(Function *F, Function *G) {
1562 if (!G->mayBeOverridden()) {
1563 // Redirect direct callers of G to F.
1564 replaceDirectCallers(G, F);
1567 // If G was internal then we may have replaced all uses of G with F. If so,
1568 // stop here and delete G. There's no need for a thunk.
1569 if (G->hasLocalLinkage() && G->use_empty()) {
1570 G->eraseFromParent();
1574 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
1576 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
1577 IRBuilder<false> Builder(BB);
1579 SmallVector<Value *, 16> Args;
1581 FunctionType *FFTy = F->getFunctionType();
1582 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
1584 Args.push_back(createCast(Builder, (Value*)AI, FFTy->getParamType(i)));
1588 CallInst *CI = Builder.CreateCall(F, Args);
1590 CI->setCallingConv(F->getCallingConv());
1591 if (NewG->getReturnType()->isVoidTy()) {
1592 Builder.CreateRetVoid();
1594 Builder.CreateRet(createCast(Builder, CI, NewG->getReturnType()));
1597 NewG->copyAttributesFrom(G);
1600 G->replaceAllUsesWith(NewG);
1601 G->eraseFromParent();
1603 DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n');
1607 // Replace G with an alias to F and delete G.
1608 void MergeFunctions::writeAlias(Function *F, Function *G) {
1609 PointerType *PTy = G->getType();
1610 auto *GA = GlobalAlias::create(PTy, G->getLinkage(), "", F);
1611 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
1613 GA->setVisibility(G->getVisibility());
1615 G->replaceAllUsesWith(GA);
1616 G->eraseFromParent();
1618 DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n');
1619 ++NumAliasesWritten;
1622 // Merge two equivalent functions. Upon completion, Function G is deleted.
1623 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) {
1624 if (F->mayBeOverridden()) {
1625 assert(G->mayBeOverridden());
1627 // Make them both thunks to the same internal function.
1628 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
1630 H->copyAttributesFrom(F);
1633 F->replaceAllUsesWith(H);
1635 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
1637 if (HasGlobalAliases) {
1645 F->setAlignment(MaxAlignment);
1646 F->setLinkage(GlobalValue::PrivateLinkage);
1649 writeThunkOrAlias(F, G);
1652 ++NumFunctionsMerged;
1655 /// Replace function F for function G in the map.
1656 void MergeFunctions::replaceFunctionInTree(FnTreeType::iterator &IterToF,
1658 Function *F = IterToF->getFunc();
1660 // A total order is already guaranteed otherwise because we process strong
1661 // functions before weak functions.
1662 assert(((F->mayBeOverridden() && G->mayBeOverridden()) ||
1663 (!F->mayBeOverridden() && !G->mayBeOverridden())) &&
1664 "Only change functions if both are strong or both are weak");
1666 assert(FunctionComparator(F, G, &GlobalNumbers).compare() == 0 &&
1667 "The two functions must be equal");
1669 IterToF->replaceBy(G);
1672 // Insert a ComparableFunction into the FnTree, or merge it away if equal to one
1673 // that was already inserted.
1674 bool MergeFunctions::insert(Function *NewFunction) {
1675 std::pair<FnTreeType::iterator, bool> Result =
1676 FnTree.insert(FunctionNode(NewFunction));
1678 if (Result.second) {
1679 DEBUG(dbgs() << "Inserting as unique: " << NewFunction->getName() << '\n');
1683 const FunctionNode &OldF = *Result.first;
1685 // Don't merge tiny functions, since it can just end up making the function
1687 // FIXME: Should still merge them if they are unnamed_addr and produce an
1689 if (NewFunction->size() == 1) {
1690 if (NewFunction->front().size() <= 2) {
1691 DEBUG(dbgs() << NewFunction->getName()
1692 << " is to small to bother merging\n");
1697 // Impose a total order (by name) on the replacement of functions. This is
1698 // important when operating on more than one module independently to prevent
1699 // cycles of thunks calling each other when the modules are linked together.
1701 // When one function is weak and the other is strong there is an order imposed
1702 // already. We process strong functions before weak functions.
1703 if ((OldF.getFunc()->mayBeOverridden() && NewFunction->mayBeOverridden()) ||
1704 (!OldF.getFunc()->mayBeOverridden() && !NewFunction->mayBeOverridden()))
1705 if (OldF.getFunc()->getName() > NewFunction->getName()) {
1706 // Swap the two functions.
1707 Function *F = OldF.getFunc();
1708 replaceFunctionInTree(Result.first, NewFunction);
1710 assert(OldF.getFunc() != F && "Must have swapped the functions.");
1713 // Never thunk a strong function to a weak function.
1714 assert(!OldF.getFunc()->mayBeOverridden() || NewFunction->mayBeOverridden());
1716 DEBUG(dbgs() << " " << OldF.getFunc()->getName()
1717 << " == " << NewFunction->getName() << '\n');
1719 Function *DeleteF = NewFunction;
1720 mergeTwoFunctions(OldF.getFunc(), DeleteF);
1724 // Remove a function from FnTree. If it was already in FnTree, add
1725 // it to Deferred so that we'll look at it in the next round.
1726 void MergeFunctions::remove(Function *F) {
1727 // We need to make sure we remove F, not a function "equal" to F per the
1728 // function equality comparator.
1729 FnTreeType::iterator found = FnTree.find(FunctionNode(F));
1731 if (found != FnTree.end() && found->getFunc() == F) {
1733 FnTree.erase(found);
1737 DEBUG(dbgs() << "Removed " << F->getName()
1738 << " from set and deferred it.\n");
1739 Deferred.emplace_back(F);
1743 // For each instruction used by the value, remove() the function that contains
1744 // the instruction. This should happen right before a call to RAUW.
1745 void MergeFunctions::removeUsers(Value *V) {
1746 std::vector<Value *> Worklist;
1747 Worklist.push_back(V);
1748 SmallSet<Value*, 8> Visited;
1750 while (!Worklist.empty()) {
1751 Value *V = Worklist.back();
1752 Worklist.pop_back();
1754 for (User *U : V->users()) {
1755 if (Instruction *I = dyn_cast<Instruction>(U)) {
1756 remove(I->getParent()->getParent());
1757 } else if (isa<GlobalValue>(U)) {
1759 } else if (Constant *C = dyn_cast<Constant>(U)) {
1760 for (User *UU : C->users()) {
1761 if (!Visited.insert(UU).second)
1762 Worklist.push_back(UU);