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"
117 using namespace llvm;
119 #define DEBUG_TYPE "mergefunc"
121 STATISTIC(NumFunctionsMerged, "Number of functions merged");
122 STATISTIC(NumThunksWritten, "Number of thunks generated");
123 STATISTIC(NumAliasesWritten, "Number of aliases generated");
124 STATISTIC(NumDoubleWeak, "Number of new functions created");
126 static cl::opt<unsigned> NumFunctionsForSanityCheck(
128 cl::desc("How many functions in module could be used for "
129 "MergeFunctions pass sanity check. "
130 "'0' disables this check. Works only with '-debug' key."),
131 cl::init(0), cl::Hidden);
135 /// GlobalNumberState assigns an integer to each global value in the program,
136 /// which is used by the comparison routine to order references to globals. This
137 /// state must be preserved throughout the pass, because Functions and other
138 /// globals need to maintain their relative order. Globals are assigned a number
139 /// when they are first visited. This order is deterministic, and so the
140 /// assigned numbers are as well. When two functions are merged, neither number
141 /// is updated. If the symbols are weak, this would be incorrect. If they are
142 /// strong, then one will be replaced at all references to the other, and so
143 /// direct callsites will now see one or the other symbol, and no update is
144 /// necessary. Note that if we were guaranteed unique names, we could just
145 /// compare those, but this would not work for stripped bitcodes or for those
146 /// few symbols without a name.
147 class GlobalNumberState {
148 struct Config : ValueMapConfig<GlobalValue*> {
149 enum { FollowRAUW = false };
151 // Each GlobalValue is mapped to an identifier. The Config ensures when RAUW
152 // occurs, the mapping does not change. Tracking changes is unnecessary, and
153 // also problematic for weak symbols (which may be overwritten).
154 typedef ValueMap<GlobalValue *, uint64_t, Config> ValueNumberMap;
155 ValueNumberMap GlobalNumbers;
156 // The next unused serial number to assign to a global.
159 GlobalNumberState() : GlobalNumbers(), NextNumber(0) {}
160 uint64_t getNumber(GlobalValue* Global) {
161 ValueNumberMap::iterator MapIter;
163 std::tie(MapIter, Inserted) = GlobalNumbers.insert({Global, NextNumber});
166 return MapIter->second;
169 GlobalNumbers.clear();
173 /// FunctionComparator - Compares two functions to determine whether or not
174 /// they will generate machine code with the same behaviour. DataLayout is
175 /// used if available. The comparator always fails conservatively (erring on the
176 /// side of claiming that two functions are different).
177 class FunctionComparator {
179 FunctionComparator(const Function *F1, const Function *F2,
180 GlobalNumberState* GN)
181 : FnL(F1), FnR(F2), GlobalNumbers(GN) {}
183 /// Test whether the two functions have equivalent behaviour.
185 /// Hash a function. Equivalent functions will have the same hash, and unequal
186 /// functions will have different hashes with high probability.
187 typedef uint64_t FunctionHash;
188 static FunctionHash functionHash(Function &);
191 /// Test whether two basic blocks have equivalent behaviour.
192 int cmpBasicBlocks(const BasicBlock *BBL, const BasicBlock *BBR);
194 /// Constants comparison.
195 /// Its analog to lexicographical comparison between hypothetical numbers
197 /// <bitcastability-trait><raw-bit-contents>
199 /// 1. Bitcastability.
200 /// Check whether L's type could be losslessly bitcasted to R's type.
201 /// On this stage method, in case when lossless bitcast is not possible
202 /// method returns -1 or 1, thus also defining which type is greater in
203 /// context of bitcastability.
204 /// Stage 0: If types are equal in terms of cmpTypes, then we can go straight
205 /// to the contents comparison.
206 /// If types differ, remember types comparison result and check
207 /// whether we still can bitcast types.
208 /// Stage 1: Types that satisfies isFirstClassType conditions are always
209 /// greater then others.
210 /// Stage 2: Vector is greater then non-vector.
211 /// If both types are vectors, then vector with greater bitwidth is
213 /// If both types are vectors with the same bitwidth, then types
214 /// are bitcastable, and we can skip other stages, and go to contents
216 /// Stage 3: Pointer types are greater than non-pointers. If both types are
217 /// pointers of the same address space - go to contents comparison.
218 /// Different address spaces: pointer with greater address space is
220 /// Stage 4: Types are neither vectors, nor pointers. And they differ.
221 /// We don't know how to bitcast them. So, we better don't do it,
222 /// and return types comparison result (so it determines the
223 /// relationship among constants we don't know how to bitcast).
225 /// Just for clearance, let's see how the set of constants could look
226 /// on single dimension axis:
228 /// [NFCT], [FCT, "others"], [FCT, pointers], [FCT, vectors]
229 /// Where: NFCT - Not a FirstClassType
230 /// FCT - FirstClassTyp:
232 /// 2. Compare raw contents.
233 /// It ignores types on this stage and only compares bits from L and R.
234 /// Returns 0, if L and R has equivalent contents.
235 /// -1 or 1 if values are different.
237 /// 2.1. If contents are numbers, compare numbers.
238 /// Ints with greater bitwidth are greater. Ints with same bitwidths
239 /// compared by their contents.
240 /// 2.2. "And so on". Just to avoid discrepancies with comments
241 /// perhaps it would be better to read the implementation itself.
242 /// 3. And again about overall picture. Let's look back at how the ordered set
243 /// of constants will look like:
244 /// [NFCT], [FCT, "others"], [FCT, pointers], [FCT, vectors]
246 /// Now look, what could be inside [FCT, "others"], for example:
247 /// [FCT, "others"] =
249 /// [double 0.1], [double 1.23],
250 /// [i32 1], [i32 2],
251 /// { double 1.0 }, ; StructTyID, NumElements = 1
252 /// { i32 1 }, ; StructTyID, NumElements = 1
253 /// { double 1, i32 1 }, ; StructTyID, NumElements = 2
254 /// { i32 1, double 1 } ; StructTyID, NumElements = 2
257 /// Let's explain the order. Float numbers will be less than integers, just
258 /// because of cmpType terms: FloatTyID < IntegerTyID.
259 /// Floats (with same fltSemantics) are sorted according to their value.
260 /// Then you can see integers, and they are, like a floats,
261 /// could be easy sorted among each others.
262 /// The structures. Structures are grouped at the tail, again because of their
263 /// TypeID: StructTyID > IntegerTyID > FloatTyID.
264 /// Structures with greater number of elements are greater. Structures with
265 /// greater elements going first are greater.
266 /// The same logic with vectors, arrays and other possible complex types.
268 /// Bitcastable constants.
269 /// Let's assume, that some constant, belongs to some group of
270 /// "so-called-equal" values with different types, and at the same time
271 /// belongs to another group of constants with equal types
272 /// and "really" equal values.
274 /// Now, prove that this is impossible:
276 /// If constant A with type TyA is bitcastable to B with type TyB, then:
277 /// 1. All constants with equal types to TyA, are bitcastable to B. Since
278 /// those should be vectors (if TyA is vector), pointers
279 /// (if TyA is pointer), or else (if TyA equal to TyB), those types should
281 /// 2. All constants with non-equal, but bitcastable types to TyA, are
282 /// bitcastable to B.
283 /// Once again, just because we allow it to vectors and pointers only.
284 /// This statement could be expanded as below:
285 /// 2.1. All vectors with equal bitwidth to vector A, has equal bitwidth to
286 /// vector B, and thus bitcastable to B as well.
287 /// 2.2. All pointers of the same address space, no matter what they point to,
288 /// bitcastable. So if C is pointer, it could be bitcasted to A and to B.
289 /// So any constant equal or bitcastable to A is equal or bitcastable to B.
292 /// In another words, for pointers and vectors, we ignore top-level type and
293 /// look at their particular properties (bit-width for vectors, and
294 /// address space for pointers).
295 /// If these properties are equal - compare their contents.
296 int cmpConstants(const Constant *L, const Constant *R);
298 /// Compares two global values by number. Uses the GlobalNumbersState to
299 /// identify the same gobals across function calls.
300 int cmpGlobalValues(GlobalValue *L, GlobalValue *R);
302 /// Assign or look up previously assigned numbers for the two values, and
303 /// return whether the numbers are equal. Numbers are assigned in the order
305 /// Comparison order:
306 /// Stage 0: Value that is function itself is always greater then others.
307 /// If left and right values are references to their functions, then
309 /// Stage 1: Constants are greater than non-constants.
310 /// If both left and right are constants, then the result of
311 /// cmpConstants is used as cmpValues result.
312 /// Stage 2: InlineAsm instances are greater than others. If both left and
313 /// right are InlineAsm instances, InlineAsm* pointers casted to
314 /// integers and compared as numbers.
315 /// Stage 3: For all other cases we compare order we meet these values in
316 /// their functions. If right value was met first during scanning,
317 /// then left value is greater.
318 /// In another words, we compare serial numbers, for more details
319 /// see comments for sn_mapL and sn_mapR.
320 int cmpValues(const Value *L, const Value *R);
322 /// Compare two Instructions for equivalence, similar to
323 /// Instruction::isSameOperationAs but with modifications to the type
325 /// Stages are listed in "most significant stage first" order:
326 /// On each stage below, we do comparison between some left and right
327 /// operation parts. If parts are non-equal, we assign parts comparison
328 /// result to the operation comparison result and exit from method.
329 /// Otherwise we proceed to the next stage.
331 /// 1. Operations opcodes. Compared as numbers.
332 /// 2. Number of operands.
333 /// 3. Operation types. Compared with cmpType method.
334 /// 4. Compare operation subclass optional data as stream of bytes:
335 /// just convert it to integers and call cmpNumbers.
336 /// 5. Compare in operation operand types with cmpType in
337 /// most significant operand first order.
338 /// 6. Last stage. Check operations for some specific attributes.
339 /// For example, for Load it would be:
340 /// 6.1.Load: volatile (as boolean flag)
341 /// 6.2.Load: alignment (as integer numbers)
342 /// 6.3.Load: synch-scope (as integer numbers)
343 /// 6.4.Load: range metadata (as integer numbers)
344 /// On this stage its better to see the code, since its not more than 10-15
345 /// strings for particular instruction, and could change sometimes.
346 int cmpOperations(const Instruction *L, const Instruction *R) const;
348 /// Compare two GEPs for equivalent pointer arithmetic.
349 /// Parts to be compared for each comparison stage,
350 /// most significant stage first:
351 /// 1. Address space. As numbers.
352 /// 2. Constant offset, (using GEPOperator::accumulateConstantOffset method).
353 /// 3. Pointer operand type (using cmpType method).
354 /// 4. Number of operands.
355 /// 5. Compare operands, using cmpValues method.
356 int cmpGEPs(const GEPOperator *GEPL, const GEPOperator *GEPR);
357 int cmpGEPs(const GetElementPtrInst *GEPL, const GetElementPtrInst *GEPR) {
358 return cmpGEPs(cast<GEPOperator>(GEPL), cast<GEPOperator>(GEPR));
361 /// cmpType - compares two types,
362 /// defines total ordering among the types set.
365 /// 0 if types are equal,
366 /// -1 if Left is less than Right,
367 /// +1 if Left is greater than Right.
370 /// Comparison is broken onto stages. Like in lexicographical comparison
371 /// stage coming first has higher priority.
372 /// On each explanation stage keep in mind total ordering properties.
374 /// 0. Before comparison we coerce pointer types of 0 address space to
376 /// We also don't bother with same type at left and right, so
377 /// just return 0 in this case.
379 /// 1. If types are of different kind (different type IDs).
380 /// Return result of type IDs comparison, treating them as numbers.
381 /// 2. If types are integers, check that they have the same width. If they
382 /// are vectors, check that they have the same count and subtype.
383 /// 3. Types have the same ID, so check whether they are one of:
392 /// We can treat these types as equal whenever their IDs are same.
393 /// 4. If Left and Right are pointers, return result of address space
394 /// comparison (numbers comparison). We can treat pointer types of same
395 /// address space as equal.
396 /// 5. If types are complex.
397 /// Then both Left and Right are to be expanded and their element types will
398 /// be checked with the same way. If we get Res != 0 on some stage, return it.
399 /// Otherwise return 0.
400 /// 6. For all other cases put llvm_unreachable.
401 int cmpTypes(Type *TyL, Type *TyR) const;
403 int cmpNumbers(uint64_t L, uint64_t R) const;
404 int cmpAPInts(const APInt &L, const APInt &R) const;
405 int cmpAPFloats(const APFloat &L, const APFloat &R) const;
406 int cmpInlineAsm(const InlineAsm *L, const InlineAsm *R) const;
407 int cmpMem(StringRef L, StringRef R) const;
408 int cmpAttrs(const AttributeSet L, const AttributeSet R) const;
409 int cmpRangeMetadata(const MDNode* L, const MDNode* R) const;
411 // The two functions undergoing comparison.
412 const Function *FnL, *FnR;
414 /// Assign serial numbers to values from left function, and values from
417 /// Being comparing functions we need to compare values we meet at left and
419 /// Its easy to sort things out for external values. It just should be
420 /// the same value at left and right.
421 /// But for local values (those were introduced inside function body)
422 /// we have to ensure they were introduced at exactly the same place,
423 /// and plays the same role.
424 /// Let's assign serial number to each value when we meet it first time.
425 /// Values that were met at same place will be with same serial numbers.
426 /// In this case it would be good to explain few points about values assigned
427 /// to BBs and other ways of implementation (see below).
429 /// 1. Safety of BB reordering.
430 /// It's safe to change the order of BasicBlocks in function.
431 /// Relationship with other functions and serial numbering will not be
432 /// changed in this case.
433 /// As follows from FunctionComparator::compare(), we do CFG walk: we start
434 /// from the entry, and then take each terminator. So it doesn't matter how in
435 /// fact BBs are ordered in function. And since cmpValues are called during
436 /// this walk, the numbering depends only on how BBs located inside the CFG.
437 /// So the answer is - yes. We will get the same numbering.
439 /// 2. Impossibility to use dominance properties of values.
440 /// If we compare two instruction operands: first is usage of local
441 /// variable AL from function FL, and second is usage of local variable AR
442 /// from FR, we could compare their origins and check whether they are
443 /// defined at the same place.
444 /// But, we are still not able to compare operands of PHI nodes, since those
445 /// could be operands from further BBs we didn't scan yet.
446 /// So it's impossible to use dominance properties in general.
447 DenseMap<const Value*, int> sn_mapL, sn_mapR;
449 // The global state we will use
450 GlobalNumberState* GlobalNumbers;
454 mutable AssertingVH<Function> F;
455 FunctionComparator::FunctionHash Hash;
457 // Note the hash is recalculated potentially multiple times, but it is cheap.
458 FunctionNode(Function *F)
459 : F(F), Hash(FunctionComparator::functionHash(*F)) {}
460 Function *getFunc() const { return F; }
461 FunctionComparator::FunctionHash getHash() const { return Hash; }
463 /// Replace the reference to the function F by the function G, assuming their
464 /// implementations are equal.
465 void replaceBy(Function *G) const {
469 void release() { F = nullptr; }
471 } // end anonymous namespace
473 int FunctionComparator::cmpNumbers(uint64_t L, uint64_t R) const {
474 if (L < R) return -1;
479 int FunctionComparator::cmpAPInts(const APInt &L, const APInt &R) const {
480 if (int Res = cmpNumbers(L.getBitWidth(), R.getBitWidth()))
482 if (L.ugt(R)) return 1;
483 if (R.ugt(L)) return -1;
487 int FunctionComparator::cmpAPFloats(const APFloat &L, const APFloat &R) const {
488 // Floats are ordered first by semantics (i.e. float, double, half, etc.),
489 // then by value interpreted as a bitstring (aka APInt).
490 const fltSemantics &SL = L.getSemantics(), &SR = R.getSemantics();
491 if (int Res = cmpNumbers(APFloat::semanticsPrecision(SL),
492 APFloat::semanticsPrecision(SR)))
494 if (int Res = cmpNumbers(APFloat::semanticsMaxExponent(SL),
495 APFloat::semanticsMaxExponent(SR)))
497 if (int Res = cmpNumbers(APFloat::semanticsMinExponent(SL),
498 APFloat::semanticsMinExponent(SR)))
500 if (int Res = cmpNumbers(APFloat::semanticsSizeInBits(SL),
501 APFloat::semanticsSizeInBits(SR)))
503 return cmpAPInts(L.bitcastToAPInt(), R.bitcastToAPInt());
506 int FunctionComparator::cmpMem(StringRef L, StringRef R) const {
507 // Prevent heavy comparison, compare sizes first.
508 if (int Res = cmpNumbers(L.size(), R.size()))
511 // Compare strings lexicographically only when it is necessary: only when
512 // strings are equal in size.
516 int FunctionComparator::cmpAttrs(const AttributeSet L,
517 const AttributeSet R) const {
518 if (int Res = cmpNumbers(L.getNumSlots(), R.getNumSlots()))
521 for (unsigned i = 0, e = L.getNumSlots(); i != e; ++i) {
522 AttributeSet::iterator LI = L.begin(i), LE = L.end(i), RI = R.begin(i),
524 for (; LI != LE && RI != RE; ++LI, ++RI) {
540 int FunctionComparator::cmpRangeMetadata(const MDNode* L,
541 const MDNode* R) const {
548 // Range metadata is a sequence of numbers. Make sure they are the same
550 // TODO: Note that as this is metadata, it is possible to drop and/or merge
551 // this data when considering functions to merge. Thus this comparison would
552 // return 0 (i.e. equivalent), but merging would become more complicated
553 // because the ranges would need to be unioned. It is not likely that
554 // functions differ ONLY in this metadata if they are actually the same
555 // function semantically.
556 if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
558 for (size_t I = 0; I < L->getNumOperands(); ++I) {
559 ConstantInt* LLow = mdconst::extract<ConstantInt>(L->getOperand(I));
560 ConstantInt* RLow = mdconst::extract<ConstantInt>(R->getOperand(I));
561 if (int Res = cmpAPInts(LLow->getValue(), RLow->getValue()))
567 /// Constants comparison:
568 /// 1. Check whether type of L constant could be losslessly bitcasted to R
570 /// 2. Compare constant contents.
571 /// For more details see declaration comments.
572 int FunctionComparator::cmpConstants(const Constant *L, const Constant *R) {
574 Type *TyL = L->getType();
575 Type *TyR = R->getType();
577 // Check whether types are bitcastable. This part is just re-factored
578 // Type::canLosslesslyBitCastTo method, but instead of returning true/false,
579 // we also pack into result which type is "less" for us.
580 int TypesRes = cmpTypes(TyL, TyR);
582 // Types are different, but check whether we can bitcast them.
583 if (!TyL->isFirstClassType()) {
584 if (TyR->isFirstClassType())
586 // Neither TyL nor TyR are values of first class type. Return the result
587 // of comparing the types
590 if (!TyR->isFirstClassType()) {
591 if (TyL->isFirstClassType())
596 // Vector -> Vector conversions are always lossless if the two vector types
597 // have the same size, otherwise not.
598 unsigned TyLWidth = 0;
599 unsigned TyRWidth = 0;
601 if (auto *VecTyL = dyn_cast<VectorType>(TyL))
602 TyLWidth = VecTyL->getBitWidth();
603 if (auto *VecTyR = dyn_cast<VectorType>(TyR))
604 TyRWidth = VecTyR->getBitWidth();
606 if (TyLWidth != TyRWidth)
607 return cmpNumbers(TyLWidth, TyRWidth);
609 // Zero bit-width means neither TyL nor TyR are vectors.
611 PointerType *PTyL = dyn_cast<PointerType>(TyL);
612 PointerType *PTyR = dyn_cast<PointerType>(TyR);
614 unsigned AddrSpaceL = PTyL->getAddressSpace();
615 unsigned AddrSpaceR = PTyR->getAddressSpace();
616 if (int Res = cmpNumbers(AddrSpaceL, AddrSpaceR))
624 // TyL and TyR aren't vectors, nor pointers. We don't know how to
630 // OK, types are bitcastable, now check constant contents.
632 if (L->isNullValue() && R->isNullValue())
634 if (L->isNullValue() && !R->isNullValue())
636 if (!L->isNullValue() && R->isNullValue())
639 auto GlobalValueL = const_cast<GlobalValue*>(dyn_cast<GlobalValue>(L));
640 auto GlobalValueR = const_cast<GlobalValue*>(dyn_cast<GlobalValue>(R));
641 if (GlobalValueL && GlobalValueR) {
642 return cmpGlobalValues(GlobalValueL, GlobalValueR);
645 if (int Res = cmpNumbers(L->getValueID(), R->getValueID()))
648 if (const auto *SeqL = dyn_cast<ConstantDataSequential>(L)) {
649 const auto *SeqR = cast<ConstantDataSequential>(R);
650 // This handles ConstantDataArray and ConstantDataVector. Note that we
651 // compare the two raw data arrays, which might differ depending on the host
652 // endianness. This isn't a problem though, because the endiness of a module
653 // will affect the order of the constants, but this order is the same
654 // for a given input module and host platform.
655 return cmpMem(SeqL->getRawDataValues(), SeqR->getRawDataValues());
658 switch (L->getValueID()) {
659 case Value::UndefValueVal:
660 case Value::ConstantTokenNoneVal:
662 case Value::ConstantIntVal: {
663 const APInt &LInt = cast<ConstantInt>(L)->getValue();
664 const APInt &RInt = cast<ConstantInt>(R)->getValue();
665 return cmpAPInts(LInt, RInt);
667 case Value::ConstantFPVal: {
668 const APFloat &LAPF = cast<ConstantFP>(L)->getValueAPF();
669 const APFloat &RAPF = cast<ConstantFP>(R)->getValueAPF();
670 return cmpAPFloats(LAPF, RAPF);
672 case Value::ConstantArrayVal: {
673 const ConstantArray *LA = cast<ConstantArray>(L);
674 const ConstantArray *RA = cast<ConstantArray>(R);
675 uint64_t NumElementsL = cast<ArrayType>(TyL)->getNumElements();
676 uint64_t NumElementsR = cast<ArrayType>(TyR)->getNumElements();
677 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
679 for (uint64_t i = 0; i < NumElementsL; ++i) {
680 if (int Res = cmpConstants(cast<Constant>(LA->getOperand(i)),
681 cast<Constant>(RA->getOperand(i))))
686 case Value::ConstantStructVal: {
687 const ConstantStruct *LS = cast<ConstantStruct>(L);
688 const ConstantStruct *RS = cast<ConstantStruct>(R);
689 unsigned NumElementsL = cast<StructType>(TyL)->getNumElements();
690 unsigned NumElementsR = cast<StructType>(TyR)->getNumElements();
691 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
693 for (unsigned i = 0; i != NumElementsL; ++i) {
694 if (int Res = cmpConstants(cast<Constant>(LS->getOperand(i)),
695 cast<Constant>(RS->getOperand(i))))
700 case Value::ConstantVectorVal: {
701 const ConstantVector *LV = cast<ConstantVector>(L);
702 const ConstantVector *RV = cast<ConstantVector>(R);
703 unsigned NumElementsL = cast<VectorType>(TyL)->getNumElements();
704 unsigned NumElementsR = cast<VectorType>(TyR)->getNumElements();
705 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
707 for (uint64_t i = 0; i < NumElementsL; ++i) {
708 if (int Res = cmpConstants(cast<Constant>(LV->getOperand(i)),
709 cast<Constant>(RV->getOperand(i))))
714 case Value::ConstantExprVal: {
715 const ConstantExpr *LE = cast<ConstantExpr>(L);
716 const ConstantExpr *RE = cast<ConstantExpr>(R);
717 unsigned NumOperandsL = LE->getNumOperands();
718 unsigned NumOperandsR = RE->getNumOperands();
719 if (int Res = cmpNumbers(NumOperandsL, NumOperandsR))
721 for (unsigned i = 0; i < NumOperandsL; ++i) {
722 if (int Res = cmpConstants(cast<Constant>(LE->getOperand(i)),
723 cast<Constant>(RE->getOperand(i))))
728 case Value::BlockAddressVal: {
729 const BlockAddress *LBA = cast<BlockAddress>(L);
730 const BlockAddress *RBA = cast<BlockAddress>(R);
731 if (int Res = cmpValues(LBA->getFunction(), RBA->getFunction()))
733 if (LBA->getFunction() == RBA->getFunction()) {
734 // They are BBs in the same function. Order by which comes first in the
735 // BB order of the function. This order is deterministic.
736 Function* F = LBA->getFunction();
737 BasicBlock *LBB = LBA->getBasicBlock();
738 BasicBlock *RBB = RBA->getBasicBlock();
741 for(BasicBlock &BB : F->getBasicBlockList()) {
749 llvm_unreachable("Basic Block Address does not point to a basic block in "
753 // cmpValues said the functions are the same. So because they aren't
754 // literally the same pointer, they must respectively be the left and
756 assert(LBA->getFunction() == FnL && RBA->getFunction() == FnR);
757 // cmpValues will tell us if these are equivalent BasicBlocks, in the
758 // context of their respective functions.
759 return cmpValues(LBA->getBasicBlock(), RBA->getBasicBlock());
762 default: // Unknown constant, abort.
763 DEBUG(dbgs() << "Looking at valueID " << L->getValueID() << "\n");
764 llvm_unreachable("Constant ValueID not recognized.");
769 int FunctionComparator::cmpGlobalValues(GlobalValue *L, GlobalValue* R) {
770 return cmpNumbers(GlobalNumbers->getNumber(L), GlobalNumbers->getNumber(R));
773 /// cmpType - compares two types,
774 /// defines total ordering among the types set.
775 /// See method declaration comments for more details.
776 int FunctionComparator::cmpTypes(Type *TyL, Type *TyR) const {
777 PointerType *PTyL = dyn_cast<PointerType>(TyL);
778 PointerType *PTyR = dyn_cast<PointerType>(TyR);
780 const DataLayout &DL = FnL->getParent()->getDataLayout();
781 if (PTyL && PTyL->getAddressSpace() == 0)
782 TyL = DL.getIntPtrType(TyL);
783 if (PTyR && PTyR->getAddressSpace() == 0)
784 TyR = DL.getIntPtrType(TyR);
789 if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
792 switch (TyL->getTypeID()) {
794 llvm_unreachable("Unknown type!");
795 // Fall through in Release mode.
796 case Type::IntegerTyID:
797 return cmpNumbers(cast<IntegerType>(TyL)->getBitWidth(),
798 cast<IntegerType>(TyR)->getBitWidth());
799 case Type::VectorTyID: {
800 VectorType *VTyL = cast<VectorType>(TyL), *VTyR = cast<VectorType>(TyR);
801 if (int Res = cmpNumbers(VTyL->getNumElements(), VTyR->getNumElements()))
803 return cmpTypes(VTyL->getElementType(), VTyR->getElementType());
805 // TyL == TyR would have returned true earlier, because types are uniqued.
807 case Type::FloatTyID:
808 case Type::DoubleTyID:
809 case Type::X86_FP80TyID:
810 case Type::FP128TyID:
811 case Type::PPC_FP128TyID:
812 case Type::LabelTyID:
813 case Type::MetadataTyID:
814 case Type::TokenTyID:
817 case Type::PointerTyID: {
818 assert(PTyL && PTyR && "Both types must be pointers here.");
819 return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
822 case Type::StructTyID: {
823 StructType *STyL = cast<StructType>(TyL);
824 StructType *STyR = cast<StructType>(TyR);
825 if (STyL->getNumElements() != STyR->getNumElements())
826 return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
828 if (STyL->isPacked() != STyR->isPacked())
829 return cmpNumbers(STyL->isPacked(), STyR->isPacked());
831 for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
832 if (int Res = cmpTypes(STyL->getElementType(i), STyR->getElementType(i)))
838 case Type::FunctionTyID: {
839 FunctionType *FTyL = cast<FunctionType>(TyL);
840 FunctionType *FTyR = cast<FunctionType>(TyR);
841 if (FTyL->getNumParams() != FTyR->getNumParams())
842 return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
844 if (FTyL->isVarArg() != FTyR->isVarArg())
845 return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
847 if (int Res = cmpTypes(FTyL->getReturnType(), FTyR->getReturnType()))
850 for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
851 if (int Res = cmpTypes(FTyL->getParamType(i), FTyR->getParamType(i)))
857 case Type::ArrayTyID: {
858 ArrayType *ATyL = cast<ArrayType>(TyL);
859 ArrayType *ATyR = cast<ArrayType>(TyR);
860 if (ATyL->getNumElements() != ATyR->getNumElements())
861 return cmpNumbers(ATyL->getNumElements(), ATyR->getNumElements());
862 return cmpTypes(ATyL->getElementType(), ATyR->getElementType());
867 // Determine whether the two operations are the same except that pointer-to-A
868 // and pointer-to-B are equivalent. This should be kept in sync with
869 // Instruction::isSameOperationAs.
870 // Read method declaration comments for more details.
871 int FunctionComparator::cmpOperations(const Instruction *L,
872 const Instruction *R) const {
873 // Differences from Instruction::isSameOperationAs:
874 // * replace type comparison with calls to isEquivalentType.
875 // * we test for I->hasSameSubclassOptionalData (nuw/nsw/tail) at the top
876 // * because of the above, we don't test for the tail bit on calls later on
877 if (int Res = cmpNumbers(L->getOpcode(), R->getOpcode()))
880 if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
883 if (int Res = cmpTypes(L->getType(), R->getType()))
886 if (int Res = cmpNumbers(L->getRawSubclassOptionalData(),
887 R->getRawSubclassOptionalData()))
890 if (const AllocaInst *AI = dyn_cast<AllocaInst>(L)) {
891 if (int Res = cmpTypes(AI->getAllocatedType(),
892 cast<AllocaInst>(R)->getAllocatedType()))
895 cmpNumbers(AI->getAlignment(), cast<AllocaInst>(R)->getAlignment()))
899 // We have two instructions of identical opcode and #operands. Check to see
900 // if all operands are the same type
901 for (unsigned i = 0, e = L->getNumOperands(); i != e; ++i) {
903 cmpTypes(L->getOperand(i)->getType(), R->getOperand(i)->getType()))
907 // Check special state that is a part of some instructions.
908 if (const LoadInst *LI = dyn_cast<LoadInst>(L)) {
909 if (int Res = cmpNumbers(LI->isVolatile(), cast<LoadInst>(R)->isVolatile()))
912 cmpNumbers(LI->getAlignment(), cast<LoadInst>(R)->getAlignment()))
915 cmpNumbers(LI->getOrdering(), cast<LoadInst>(R)->getOrdering()))
918 cmpNumbers(LI->getSynchScope(), cast<LoadInst>(R)->getSynchScope()))
920 return cmpRangeMetadata(LI->getMetadata(LLVMContext::MD_range),
921 cast<LoadInst>(R)->getMetadata(LLVMContext::MD_range));
923 if (const StoreInst *SI = dyn_cast<StoreInst>(L)) {
925 cmpNumbers(SI->isVolatile(), cast<StoreInst>(R)->isVolatile()))
928 cmpNumbers(SI->getAlignment(), cast<StoreInst>(R)->getAlignment()))
931 cmpNumbers(SI->getOrdering(), cast<StoreInst>(R)->getOrdering()))
933 return cmpNumbers(SI->getSynchScope(), cast<StoreInst>(R)->getSynchScope());
935 if (const CmpInst *CI = dyn_cast<CmpInst>(L))
936 return cmpNumbers(CI->getPredicate(), cast<CmpInst>(R)->getPredicate());
937 if (const CallInst *CI = dyn_cast<CallInst>(L)) {
938 if (int Res = cmpNumbers(CI->getCallingConv(),
939 cast<CallInst>(R)->getCallingConv()))
942 cmpAttrs(CI->getAttributes(), cast<CallInst>(R)->getAttributes()))
944 return cmpRangeMetadata(
945 CI->getMetadata(LLVMContext::MD_range),
946 cast<CallInst>(R)->getMetadata(LLVMContext::MD_range));
948 if (const InvokeInst *CI = dyn_cast<InvokeInst>(L)) {
949 if (int Res = cmpNumbers(CI->getCallingConv(),
950 cast<InvokeInst>(R)->getCallingConv()))
953 cmpAttrs(CI->getAttributes(), cast<InvokeInst>(R)->getAttributes()))
955 return cmpRangeMetadata(
956 CI->getMetadata(LLVMContext::MD_range),
957 cast<InvokeInst>(R)->getMetadata(LLVMContext::MD_range));
959 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(L)) {
960 ArrayRef<unsigned> LIndices = IVI->getIndices();
961 ArrayRef<unsigned> RIndices = cast<InsertValueInst>(R)->getIndices();
962 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
964 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
965 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
969 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(L)) {
970 ArrayRef<unsigned> LIndices = EVI->getIndices();
971 ArrayRef<unsigned> RIndices = cast<ExtractValueInst>(R)->getIndices();
972 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
974 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
975 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
979 if (const FenceInst *FI = dyn_cast<FenceInst>(L)) {
981 cmpNumbers(FI->getOrdering(), cast<FenceInst>(R)->getOrdering()))
983 return cmpNumbers(FI->getSynchScope(), cast<FenceInst>(R)->getSynchScope());
986 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(L)) {
987 if (int Res = cmpNumbers(CXI->isVolatile(),
988 cast<AtomicCmpXchgInst>(R)->isVolatile()))
990 if (int Res = cmpNumbers(CXI->isWeak(),
991 cast<AtomicCmpXchgInst>(R)->isWeak()))
993 if (int Res = cmpNumbers(CXI->getSuccessOrdering(),
994 cast<AtomicCmpXchgInst>(R)->getSuccessOrdering()))
996 if (int Res = cmpNumbers(CXI->getFailureOrdering(),
997 cast<AtomicCmpXchgInst>(R)->getFailureOrdering()))
999 return cmpNumbers(CXI->getSynchScope(),
1000 cast<AtomicCmpXchgInst>(R)->getSynchScope());
1002 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(L)) {
1003 if (int Res = cmpNumbers(RMWI->getOperation(),
1004 cast<AtomicRMWInst>(R)->getOperation()))
1006 if (int Res = cmpNumbers(RMWI->isVolatile(),
1007 cast<AtomicRMWInst>(R)->isVolatile()))
1009 if (int Res = cmpNumbers(RMWI->getOrdering(),
1010 cast<AtomicRMWInst>(R)->getOrdering()))
1012 return cmpNumbers(RMWI->getSynchScope(),
1013 cast<AtomicRMWInst>(R)->getSynchScope());
1018 // Determine whether two GEP operations perform the same underlying arithmetic.
1019 // Read method declaration comments for more details.
1020 int FunctionComparator::cmpGEPs(const GEPOperator *GEPL,
1021 const GEPOperator *GEPR) {
1023 unsigned int ASL = GEPL->getPointerAddressSpace();
1024 unsigned int ASR = GEPR->getPointerAddressSpace();
1026 if (int Res = cmpNumbers(ASL, ASR))
1029 // When we have target data, we can reduce the GEP down to the value in bytes
1030 // added to the address.
1031 const DataLayout &DL = FnL->getParent()->getDataLayout();
1032 unsigned BitWidth = DL.getPointerSizeInBits(ASL);
1033 APInt OffsetL(BitWidth, 0), OffsetR(BitWidth, 0);
1034 if (GEPL->accumulateConstantOffset(DL, OffsetL) &&
1035 GEPR->accumulateConstantOffset(DL, OffsetR))
1036 return cmpAPInts(OffsetL, OffsetR);
1037 if (int Res = cmpTypes(GEPL->getSourceElementType(),
1038 GEPR->getSourceElementType()))
1041 if (int Res = cmpNumbers(GEPL->getNumOperands(), GEPR->getNumOperands()))
1044 for (unsigned i = 0, e = GEPL->getNumOperands(); i != e; ++i) {
1045 if (int Res = cmpValues(GEPL->getOperand(i), GEPR->getOperand(i)))
1052 int FunctionComparator::cmpInlineAsm(const InlineAsm *L,
1053 const InlineAsm *R) const {
1054 // InlineAsm's are uniqued. If they are the same pointer, obviously they are
1055 // the same, otherwise compare the fields.
1058 if (int Res = cmpTypes(L->getFunctionType(), R->getFunctionType()))
1060 if (int Res = cmpMem(L->getAsmString(), R->getAsmString()))
1062 if (int Res = cmpMem(L->getConstraintString(), R->getConstraintString()))
1064 if (int Res = cmpNumbers(L->hasSideEffects(), R->hasSideEffects()))
1066 if (int Res = cmpNumbers(L->isAlignStack(), R->isAlignStack()))
1068 if (int Res = cmpNumbers(L->getDialect(), R->getDialect()))
1070 llvm_unreachable("InlineAsm blocks were not uniqued.");
1074 /// Compare two values used by the two functions under pair-wise comparison. If
1075 /// this is the first time the values are seen, they're added to the mapping so
1076 /// that we will detect mismatches on next use.
1077 /// See comments in declaration for more details.
1078 int FunctionComparator::cmpValues(const Value *L, const Value *R) {
1079 // Catch self-reference case.
1091 const Constant *ConstL = dyn_cast<Constant>(L);
1092 const Constant *ConstR = dyn_cast<Constant>(R);
1093 if (ConstL && ConstR) {
1096 return cmpConstants(ConstL, ConstR);
1104 const InlineAsm *InlineAsmL = dyn_cast<InlineAsm>(L);
1105 const InlineAsm *InlineAsmR = dyn_cast<InlineAsm>(R);
1107 if (InlineAsmL && InlineAsmR)
1108 return cmpInlineAsm(InlineAsmL, InlineAsmR);
1114 auto LeftSN = sn_mapL.insert(std::make_pair(L, sn_mapL.size())),
1115 RightSN = sn_mapR.insert(std::make_pair(R, sn_mapR.size()));
1117 return cmpNumbers(LeftSN.first->second, RightSN.first->second);
1119 // Test whether two basic blocks have equivalent behaviour.
1120 int FunctionComparator::cmpBasicBlocks(const BasicBlock *BBL,
1121 const BasicBlock *BBR) {
1122 BasicBlock::const_iterator InstL = BBL->begin(), InstLE = BBL->end();
1123 BasicBlock::const_iterator InstR = BBR->begin(), InstRE = BBR->end();
1126 if (int Res = cmpValues(&*InstL, &*InstR))
1129 const GetElementPtrInst *GEPL = dyn_cast<GetElementPtrInst>(InstL);
1130 const GetElementPtrInst *GEPR = dyn_cast<GetElementPtrInst>(InstR);
1139 cmpValues(GEPL->getPointerOperand(), GEPR->getPointerOperand()))
1141 if (int Res = cmpGEPs(GEPL, GEPR))
1144 if (int Res = cmpOperations(&*InstL, &*InstR))
1146 assert(InstL->getNumOperands() == InstR->getNumOperands());
1148 for (unsigned i = 0, e = InstL->getNumOperands(); i != e; ++i) {
1149 Value *OpL = InstL->getOperand(i);
1150 Value *OpR = InstR->getOperand(i);
1151 if (int Res = cmpValues(OpL, OpR))
1153 // cmpValues should ensure this is true.
1154 assert(cmpTypes(OpL->getType(), OpR->getType()) == 0);
1159 } while (InstL != InstLE && InstR != InstRE);
1161 if (InstL != InstLE && InstR == InstRE)
1163 if (InstL == InstLE && InstR != InstRE)
1168 // Test whether the two functions have equivalent behaviour.
1169 int FunctionComparator::compare() {
1173 if (int Res = cmpAttrs(FnL->getAttributes(), FnR->getAttributes()))
1176 if (int Res = cmpNumbers(FnL->hasGC(), FnR->hasGC()))
1180 if (int Res = cmpMem(FnL->getGC(), FnR->getGC()))
1184 if (int Res = cmpNumbers(FnL->hasSection(), FnR->hasSection()))
1187 if (FnL->hasSection()) {
1188 if (int Res = cmpMem(FnL->getSection(), FnR->getSection()))
1192 if (int Res = cmpNumbers(FnL->isVarArg(), FnR->isVarArg()))
1195 // TODO: if it's internal and only used in direct calls, we could handle this
1197 if (int Res = cmpNumbers(FnL->getCallingConv(), FnR->getCallingConv()))
1200 if (int Res = cmpTypes(FnL->getFunctionType(), FnR->getFunctionType()))
1203 assert(FnL->arg_size() == FnR->arg_size() &&
1204 "Identically typed functions have different numbers of args!");
1206 // Visit the arguments so that they get enumerated in the order they're
1208 for (Function::const_arg_iterator ArgLI = FnL->arg_begin(),
1209 ArgRI = FnR->arg_begin(),
1210 ArgLE = FnL->arg_end();
1211 ArgLI != ArgLE; ++ArgLI, ++ArgRI) {
1212 if (cmpValues(&*ArgLI, &*ArgRI) != 0)
1213 llvm_unreachable("Arguments repeat!");
1216 // We do a CFG-ordered walk since the actual ordering of the blocks in the
1217 // linked list is immaterial. Our walk starts at the entry block for both
1218 // functions, then takes each block from each terminator in order. As an
1219 // artifact, this also means that unreachable blocks are ignored.
1220 SmallVector<const BasicBlock *, 8> FnLBBs, FnRBBs;
1221 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
1223 FnLBBs.push_back(&FnL->getEntryBlock());
1224 FnRBBs.push_back(&FnR->getEntryBlock());
1226 VisitedBBs.insert(FnLBBs[0]);
1227 while (!FnLBBs.empty()) {
1228 const BasicBlock *BBL = FnLBBs.pop_back_val();
1229 const BasicBlock *BBR = FnRBBs.pop_back_val();
1231 if (int Res = cmpValues(BBL, BBR))
1234 if (int Res = cmpBasicBlocks(BBL, BBR))
1237 const TerminatorInst *TermL = BBL->getTerminator();
1238 const TerminatorInst *TermR = BBR->getTerminator();
1240 assert(TermL->getNumSuccessors() == TermR->getNumSuccessors());
1241 for (unsigned i = 0, e = TermL->getNumSuccessors(); i != e; ++i) {
1242 if (!VisitedBBs.insert(TermL->getSuccessor(i)).second)
1245 FnLBBs.push_back(TermL->getSuccessor(i));
1246 FnRBBs.push_back(TermR->getSuccessor(i));
1253 // Accumulate the hash of a sequence of 64-bit integers. This is similar to a
1254 // hash of a sequence of 64bit ints, but the entire input does not need to be
1255 // available at once. This interface is necessary for functionHash because it
1256 // needs to accumulate the hash as the structure of the function is traversed
1257 // without saving these values to an intermediate buffer. This form of hashing
1258 // is not often needed, as usually the object to hash is just read from a
1260 class HashAccumulator64 {
1263 // Initialize to random constant, so the state isn't zero.
1264 HashAccumulator64() { Hash = 0x6acaa36bef8325c5ULL; }
1265 void add(uint64_t V) {
1266 Hash = llvm::hashing::detail::hash_16_bytes(Hash, V);
1268 // No finishing is required, because the entire hash value is used.
1269 uint64_t getHash() { return Hash; }
1271 } // end anonymous namespace
1273 // A function hash is calculated by considering only the number of arguments and
1274 // whether a function is varargs, the order of basic blocks (given by the
1275 // successors of each basic block in depth first order), and the order of
1276 // opcodes of each instruction within each of these basic blocks. This mirrors
1277 // the strategy compare() uses to compare functions by walking the BBs in depth
1278 // first order and comparing each instruction in sequence. Because this hash
1279 // does not look at the operands, it is insensitive to things such as the
1280 // target of calls and the constants used in the function, which makes it useful
1281 // when possibly merging functions which are the same modulo constants and call
1283 FunctionComparator::FunctionHash FunctionComparator::functionHash(Function &F) {
1284 HashAccumulator64 H;
1285 H.add(F.isVarArg());
1286 H.add(F.arg_size());
1288 SmallVector<const BasicBlock *, 8> BBs;
1289 SmallSet<const BasicBlock *, 16> VisitedBBs;
1291 // Walk the blocks in the same order as FunctionComparator::cmpBasicBlocks(),
1292 // accumulating the hash of the function "structure." (BB and opcode sequence)
1293 BBs.push_back(&F.getEntryBlock());
1294 VisitedBBs.insert(BBs[0]);
1295 while (!BBs.empty()) {
1296 const BasicBlock *BB = BBs.pop_back_val();
1297 // This random value acts as a block header, as otherwise the partition of
1298 // opcodes into BBs wouldn't affect the hash, only the order of the opcodes
1300 for (auto &Inst : *BB) {
1301 H.add(Inst.getOpcode());
1303 const TerminatorInst *Term = BB->getTerminator();
1304 for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) {
1305 if (!VisitedBBs.insert(Term->getSuccessor(i)).second)
1307 BBs.push_back(Term->getSuccessor(i));
1316 /// MergeFunctions finds functions which will generate identical machine code,
1317 /// by considering all pointer types to be equivalent. Once identified,
1318 /// MergeFunctions will fold them by replacing a call to one to a call to a
1319 /// bitcast of the other.
1321 class MergeFunctions : public ModulePass {
1325 : ModulePass(ID), FnTree(FunctionNodeCmp(&GlobalNumbers)), FNodesInTree(),
1326 HasGlobalAliases(false) {
1327 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
1330 bool runOnModule(Module &M) override;
1333 // The function comparison operator is provided here so that FunctionNodes do
1334 // not need to become larger with another pointer.
1335 class FunctionNodeCmp {
1336 GlobalNumberState* GlobalNumbers;
1338 FunctionNodeCmp(GlobalNumberState* GN) : GlobalNumbers(GN) {}
1339 bool operator()(const FunctionNode &LHS, const FunctionNode &RHS) const {
1340 // Order first by hashes, then full function comparison.
1341 if (LHS.getHash() != RHS.getHash())
1342 return LHS.getHash() < RHS.getHash();
1343 FunctionComparator FCmp(LHS.getFunc(), RHS.getFunc(), GlobalNumbers);
1344 return FCmp.compare() == -1;
1347 typedef std::set<FunctionNode, FunctionNodeCmp> FnTreeType;
1349 GlobalNumberState GlobalNumbers;
1351 /// A work queue of functions that may have been modified and should be
1353 std::vector<WeakVH> Deferred;
1355 /// Checks the rules of order relation introduced among functions set.
1356 /// Returns true, if sanity check has been passed, and false if failed.
1357 bool doSanityCheck(std::vector<WeakVH> &Worklist);
1359 /// Insert a ComparableFunction into the FnTree, or merge it away if it's
1360 /// equal to one that's already present.
1361 bool insert(Function *NewFunction);
1363 /// Remove a Function from the FnTree and queue it up for a second sweep of
1365 void remove(Function *F);
1367 /// Find the functions that use this Value and remove them from FnTree and
1368 /// queue the functions.
1369 void removeUsers(Value *V);
1371 /// Replace all direct calls of Old with calls of New. Will bitcast New if
1372 /// necessary to make types match.
1373 void replaceDirectCallers(Function *Old, Function *New);
1375 /// Merge two equivalent functions. Upon completion, G may be deleted, or may
1376 /// be converted into a thunk. In either case, it should never be visited
1378 void mergeTwoFunctions(Function *F, Function *G);
1380 /// Replace G with a thunk or an alias to F. Deletes G.
1381 void writeThunkOrAlias(Function *F, Function *G);
1383 /// Replace G with a simple tail call to bitcast(F). Also replace direct uses
1384 /// of G with bitcast(F). Deletes G.
1385 void writeThunk(Function *F, Function *G);
1387 /// Replace G with an alias to F. Deletes G.
1388 void writeAlias(Function *F, Function *G);
1390 /// Replace function F with function G in the function tree.
1391 void replaceFunctionInTree(const FunctionNode &FN, Function *G);
1393 /// The set of all distinct functions. Use the insert() and remove() methods
1394 /// to modify it. The map allows efficient lookup and deferring of Functions.
1396 // Map functions to the iterators of the FunctionNode which contains them
1397 // in the FnTree. This must be updated carefully whenever the FnTree is
1398 // modified, i.e. in insert(), remove(), and replaceFunctionInTree(), to avoid
1399 // dangling iterators into FnTree. The invariant that preserves this is that
1400 // there is exactly one mapping F -> FN for each FunctionNode FN in FnTree.
1401 ValueMap<Function*, FnTreeType::iterator> FNodesInTree;
1403 /// Whether or not the target supports global aliases.
1404 bool HasGlobalAliases;
1407 } // end anonymous namespace
1409 char MergeFunctions::ID = 0;
1410 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
1412 ModulePass *llvm::createMergeFunctionsPass() {
1413 return new MergeFunctions();
1416 bool MergeFunctions::doSanityCheck(std::vector<WeakVH> &Worklist) {
1417 if (const unsigned Max = NumFunctionsForSanityCheck) {
1418 unsigned TripleNumber = 0;
1421 dbgs() << "MERGEFUNC-SANITY: Started for first " << Max << " functions.\n";
1424 for (std::vector<WeakVH>::iterator I = Worklist.begin(), E = Worklist.end();
1425 I != E && i < Max; ++I, ++i) {
1427 for (std::vector<WeakVH>::iterator J = I; J != E && j < Max; ++J, ++j) {
1428 Function *F1 = cast<Function>(*I);
1429 Function *F2 = cast<Function>(*J);
1430 int Res1 = FunctionComparator(F1, F2, &GlobalNumbers).compare();
1431 int Res2 = FunctionComparator(F2, F1, &GlobalNumbers).compare();
1433 // If F1 <= F2, then F2 >= F1, otherwise report failure.
1434 if (Res1 != -Res2) {
1435 dbgs() << "MERGEFUNC-SANITY: Non-symmetric; triple: " << TripleNumber
1446 for (std::vector<WeakVH>::iterator K = J; K != E && k < Max;
1447 ++k, ++K, ++TripleNumber) {
1451 Function *F3 = cast<Function>(*K);
1452 int Res3 = FunctionComparator(F1, F3, &GlobalNumbers).compare();
1453 int Res4 = FunctionComparator(F2, F3, &GlobalNumbers).compare();
1455 bool Transitive = true;
1457 if (Res1 != 0 && Res1 == Res4) {
1458 // F1 > F2, F2 > F3 => F1 > F3
1459 Transitive = Res3 == Res1;
1460 } else if (Res3 != 0 && Res3 == -Res4) {
1461 // F1 > F3, F3 > F2 => F1 > F2
1462 Transitive = Res3 == Res1;
1463 } else if (Res4 != 0 && -Res3 == Res4) {
1464 // F2 > F3, F3 > F1 => F2 > F1
1465 Transitive = Res4 == -Res1;
1469 dbgs() << "MERGEFUNC-SANITY: Non-transitive; triple: "
1470 << TripleNumber << "\n";
1471 dbgs() << "Res1, Res3, Res4: " << Res1 << ", " << Res3 << ", "
1482 dbgs() << "MERGEFUNC-SANITY: " << (Valid ? "Passed." : "Failed.") << "\n";
1488 bool MergeFunctions::runOnModule(Module &M) {
1489 bool Changed = false;
1491 // All functions in the module, ordered by hash. Functions with a unique
1492 // hash value are easily eliminated.
1493 std::vector<std::pair<FunctionComparator::FunctionHash, Function *>>
1495 for (Function &Func : M) {
1496 if (!Func.isDeclaration() && !Func.hasAvailableExternallyLinkage()) {
1497 HashedFuncs.push_back({FunctionComparator::functionHash(Func), &Func});
1502 HashedFuncs.begin(), HashedFuncs.end(),
1503 [](const std::pair<FunctionComparator::FunctionHash, Function *> &a,
1504 const std::pair<FunctionComparator::FunctionHash, Function *> &b) {
1505 return a.first < b.first;
1508 auto S = HashedFuncs.begin();
1509 for (auto I = HashedFuncs.begin(), IE = HashedFuncs.end(); I != IE; ++I) {
1510 // If the hash value matches the previous value or the next one, we must
1511 // consider merging it. Otherwise it is dropped and never considered again.
1512 if ((I != S && std::prev(I)->first == I->first) ||
1513 (std::next(I) != IE && std::next(I)->first == I->first) ) {
1514 Deferred.push_back(WeakVH(I->second));
1519 std::vector<WeakVH> Worklist;
1520 Deferred.swap(Worklist);
1522 DEBUG(doSanityCheck(Worklist));
1524 DEBUG(dbgs() << "size of module: " << M.size() << '\n');
1525 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
1527 // Insert only strong functions and merge them. Strong function merging
1528 // always deletes one of them.
1529 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
1530 E = Worklist.end(); I != E; ++I) {
1532 Function *F = cast<Function>(*I);
1533 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
1534 !F->mayBeOverridden()) {
1535 Changed |= insert(F);
1539 // Insert only weak functions and merge them. By doing these second we
1540 // create thunks to the strong function when possible. When two weak
1541 // functions are identical, we create a new strong function with two weak
1542 // weak thunks to it which are identical but not mergable.
1543 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
1544 E = Worklist.end(); I != E; ++I) {
1546 Function *F = cast<Function>(*I);
1547 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
1548 F->mayBeOverridden()) {
1549 Changed |= insert(F);
1552 DEBUG(dbgs() << "size of FnTree: " << FnTree.size() << '\n');
1553 } while (!Deferred.empty());
1556 GlobalNumbers.clear();
1561 // Replace direct callers of Old with New.
1562 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
1563 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
1564 for (auto UI = Old->use_begin(), UE = Old->use_end(); UI != UE;) {
1567 CallSite CS(U->getUser());
1568 if (CS && CS.isCallee(U)) {
1569 // Transfer the called function's attributes to the call site. Due to the
1570 // bitcast we will 'lose' ABI changing attributes because the 'called
1571 // function' is no longer a Function* but the bitcast. Code that looks up
1572 // the attributes from the called function will fail.
1574 // FIXME: This is not actually true, at least not anymore. The callsite
1575 // will always have the same ABI affecting attributes as the callee,
1576 // because otherwise the original input has UB. Note that Old and New
1577 // always have matching ABI, so no attributes need to be changed.
1578 // Transferring other attributes may help other optimizations, but that
1579 // should be done uniformly and not in this ad-hoc way.
1580 auto &Context = New->getContext();
1581 auto NewFuncAttrs = New->getAttributes();
1582 auto CallSiteAttrs = CS.getAttributes();
1584 CallSiteAttrs = CallSiteAttrs.addAttributes(
1585 Context, AttributeSet::ReturnIndex, NewFuncAttrs.getRetAttributes());
1587 for (unsigned argIdx = 0; argIdx < CS.arg_size(); argIdx++) {
1588 AttributeSet Attrs = NewFuncAttrs.getParamAttributes(argIdx);
1589 if (Attrs.getNumSlots())
1590 CallSiteAttrs = CallSiteAttrs.addAttributes(Context, argIdx, Attrs);
1593 CS.setAttributes(CallSiteAttrs);
1595 remove(CS.getInstruction()->getParent()->getParent());
1601 // Replace G with an alias to F if possible, or else a thunk to F. Deletes G.
1602 void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) {
1603 if (HasGlobalAliases && G->hasUnnamedAddr()) {
1604 if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
1605 G->hasWeakLinkage()) {
1614 // Helper for writeThunk,
1615 // Selects proper bitcast operation,
1616 // but a bit simpler then CastInst::getCastOpcode.
1617 static Value *createCast(IRBuilder<false> &Builder, Value *V, Type *DestTy) {
1618 Type *SrcTy = V->getType();
1619 if (SrcTy->isStructTy()) {
1620 assert(DestTy->isStructTy());
1621 assert(SrcTy->getStructNumElements() == DestTy->getStructNumElements());
1622 Value *Result = UndefValue::get(DestTy);
1623 for (unsigned int I = 0, E = SrcTy->getStructNumElements(); I < E; ++I) {
1624 Value *Element = createCast(
1625 Builder, Builder.CreateExtractValue(V, makeArrayRef(I)),
1626 DestTy->getStructElementType(I));
1629 Builder.CreateInsertValue(Result, Element, makeArrayRef(I));
1633 assert(!DestTy->isStructTy());
1634 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
1635 return Builder.CreateIntToPtr(V, DestTy);
1636 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
1637 return Builder.CreatePtrToInt(V, DestTy);
1639 return Builder.CreateBitCast(V, DestTy);
1642 // Replace G with a simple tail call to bitcast(F). Also replace direct uses
1643 // of G with bitcast(F). Deletes G.
1644 void MergeFunctions::writeThunk(Function *F, Function *G) {
1645 if (!G->mayBeOverridden()) {
1646 // Redirect direct callers of G to F.
1647 replaceDirectCallers(G, F);
1650 // If G was internal then we may have replaced all uses of G with F. If so,
1651 // stop here and delete G. There's no need for a thunk.
1652 if (G->hasLocalLinkage() && G->use_empty()) {
1653 G->eraseFromParent();
1657 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
1659 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
1660 IRBuilder<false> Builder(BB);
1662 SmallVector<Value *, 16> Args;
1664 FunctionType *FFTy = F->getFunctionType();
1665 for (Argument & AI : NewG->args()) {
1666 Args.push_back(createCast(Builder, &AI, FFTy->getParamType(i)));
1670 CallInst *CI = Builder.CreateCall(F, Args);
1672 CI->setCallingConv(F->getCallingConv());
1673 CI->setAttributes(F->getAttributes());
1674 if (NewG->getReturnType()->isVoidTy()) {
1675 Builder.CreateRetVoid();
1677 Builder.CreateRet(createCast(Builder, CI, NewG->getReturnType()));
1680 NewG->copyAttributesFrom(G);
1683 G->replaceAllUsesWith(NewG);
1684 G->eraseFromParent();
1686 DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n');
1690 // Replace G with an alias to F and delete G.
1691 void MergeFunctions::writeAlias(Function *F, Function *G) {
1692 auto *GA = GlobalAlias::create(G->getLinkage(), "", F);
1693 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
1695 GA->setVisibility(G->getVisibility());
1697 G->replaceAllUsesWith(GA);
1698 G->eraseFromParent();
1700 DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n');
1701 ++NumAliasesWritten;
1704 // Merge two equivalent functions. Upon completion, Function G is deleted.
1705 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) {
1706 if (F->mayBeOverridden()) {
1707 assert(G->mayBeOverridden());
1709 // Make them both thunks to the same internal function.
1710 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
1712 H->copyAttributesFrom(F);
1715 F->replaceAllUsesWith(H);
1717 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
1719 if (HasGlobalAliases) {
1727 F->setAlignment(MaxAlignment);
1728 F->setLinkage(GlobalValue::PrivateLinkage);
1731 writeThunkOrAlias(F, G);
1734 ++NumFunctionsMerged;
1737 /// Replace function F by function G.
1738 void MergeFunctions::replaceFunctionInTree(const FunctionNode &FN,
1740 Function *F = FN.getFunc();
1741 assert(FunctionComparator(F, G, &GlobalNumbers).compare() == 0 &&
1742 "The two functions must be equal");
1744 auto I = FNodesInTree.find(F);
1745 assert(I != FNodesInTree.end() && "F should be in FNodesInTree");
1746 assert(FNodesInTree.count(G) == 0 && "FNodesInTree should not contain G");
1748 FnTreeType::iterator IterToFNInFnTree = I->second;
1749 assert(&(*IterToFNInFnTree) == &FN && "F should map to FN in FNodesInTree.");
1750 // Remove F -> FN and insert G -> FN
1751 FNodesInTree.erase(I);
1752 FNodesInTree.insert({G, IterToFNInFnTree});
1753 // Replace F with G in FN, which is stored inside the FnTree.
1757 // Insert a ComparableFunction into the FnTree, or merge it away if equal to one
1758 // that was already inserted.
1759 bool MergeFunctions::insert(Function *NewFunction) {
1760 std::pair<FnTreeType::iterator, bool> Result =
1761 FnTree.insert(FunctionNode(NewFunction));
1763 if (Result.second) {
1764 assert(FNodesInTree.count(NewFunction) == 0);
1765 FNodesInTree.insert({NewFunction, Result.first});
1766 DEBUG(dbgs() << "Inserting as unique: " << NewFunction->getName() << '\n');
1770 const FunctionNode &OldF = *Result.first;
1772 // Don't merge tiny functions, since it can just end up making the function
1774 // FIXME: Should still merge them if they are unnamed_addr and produce an
1776 if (NewFunction->size() == 1) {
1777 if (NewFunction->front().size() <= 2) {
1778 DEBUG(dbgs() << NewFunction->getName()
1779 << " is to small to bother merging\n");
1784 // Impose a total order (by name) on the replacement of functions. This is
1785 // important when operating on more than one module independently to prevent
1786 // cycles of thunks calling each other when the modules are linked together.
1788 // When one function is weak and the other is strong there is an order imposed
1789 // already. We process strong functions before weak functions.
1790 if ((OldF.getFunc()->mayBeOverridden() && NewFunction->mayBeOverridden()) ||
1791 (!OldF.getFunc()->mayBeOverridden() && !NewFunction->mayBeOverridden()))
1792 if (OldF.getFunc()->getName() > NewFunction->getName()) {
1793 // Swap the two functions.
1794 Function *F = OldF.getFunc();
1795 replaceFunctionInTree(*Result.first, NewFunction);
1797 assert(OldF.getFunc() != F && "Must have swapped the functions.");
1800 // Never thunk a strong function to a weak function.
1801 assert(!OldF.getFunc()->mayBeOverridden() || NewFunction->mayBeOverridden());
1803 DEBUG(dbgs() << " " << OldF.getFunc()->getName()
1804 << " == " << NewFunction->getName() << '\n');
1806 Function *DeleteF = NewFunction;
1807 mergeTwoFunctions(OldF.getFunc(), DeleteF);
1811 // Remove a function from FnTree. If it was already in FnTree, add
1812 // it to Deferred so that we'll look at it in the next round.
1813 void MergeFunctions::remove(Function *F) {
1814 auto I = FNodesInTree.find(F);
1815 if (I != FNodesInTree.end()) {
1816 DEBUG(dbgs() << "Deferred " << F->getName()<< ".\n");
1817 FnTree.erase(I->second);
1818 // I->second has been invalidated, remove it from the FNodesInTree map to
1819 // preserve the invariant.
1820 FNodesInTree.erase(I);
1821 Deferred.emplace_back(F);
1825 // For each instruction used by the value, remove() the function that contains
1826 // the instruction. This should happen right before a call to RAUW.
1827 void MergeFunctions::removeUsers(Value *V) {
1828 std::vector<Value *> Worklist;
1829 Worklist.push_back(V);
1830 SmallSet<Value*, 8> Visited;
1832 while (!Worklist.empty()) {
1833 Value *V = Worklist.back();
1834 Worklist.pop_back();
1836 for (User *U : V->users()) {
1837 if (Instruction *I = dyn_cast<Instruction>(U)) {
1838 remove(I->getParent()->getParent());
1839 } else if (isa<GlobalValue>(U)) {
1841 } else if (Constant *C = dyn_cast<Constant>(U)) {
1842 for (User *UU : C->users()) {
1843 if (!Visited.insert(UU).second)
1844 Worklist.push_back(UU);