1 //===- MergeFunctions.cpp - Merge identical functions ---------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass looks for equivalent functions that are mergable and folds them.
11 //
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).
22 //
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.
30 //
31 // When a match is found the functions are folded. If both functions are
32 // overridable, we move the functionality into a new internal function and
33 // leave two overridable thunks to it.
34 //
35 //===----------------------------------------------------------------------===//
36 //
37 // Future work:
38 //
39 // * virtual functions.
40 //
41 // Many functions have their address taken by the virtual function table for
42 // the object they belong to. However, as long as it's only used for a lookup
43 // and call, this is irrelevant, and we'd like to fold such functions.
44 //
45 // * be smarter about bitcasts.
46 //
47 // In order to fold functions, we will sometimes add either bitcast instructions
48 // or bitcast constant expressions. Unfortunately, this can confound further
49 // analysis since the two functions differ where one has a bitcast and the
50 // other doesn't. We should learn to look through bitcasts.
51 //
52 // * Compare complex types with pointer types inside.
53 // * Compare cross-reference cases.
54 // * Compare complex expressions.
55 //
56 // All the three issues above could be described as ability to prove that
57 // fA == fB == fC == fE == fF == fG in example below:
58 //
59 // void fA() {
60 // fB();
61 // }
62 // void fB() {
63 // fA();
64 // }
65 //
66 // void fE() {
67 // fF();
68 // }
69 // void fF() {
70 // fG();
71 // }
72 // void fG() {
73 // fE();
74 // }
75 //
76 // Simplest cross-reference case (fA <--> fB) was implemented in previous
77 // versions of MergeFunctions, though it presented only in two function pairs
78 // in test-suite (that counts >50k functions)
79 // Though possibility to detect complex cross-referencing (e.g.: A->B->C->D->A)
80 // could cover much more cases.
81 //
82 //===----------------------------------------------------------------------===//
83
84 #include "llvm/Transforms/IPO.h"
85 #include "llvm/ADT/DenseSet.h"
86 #include "llvm/ADT/FoldingSet.h"
87 #include "llvm/ADT/STLExtras.h"
88 #include "llvm/ADT/SmallSet.h"
89 #include "llvm/ADT/Statistic.h"
90 #include "llvm/IR/CallSite.h"
91 #include "llvm/IR/Constants.h"
92 #include "llvm/IR/DataLayout.h"
93 #include "llvm/IR/IRBuilder.h"
94 #include "llvm/IR/InlineAsm.h"
95 #include "llvm/IR/Instructions.h"
96 #include "llvm/IR/LLVMContext.h"
97 #include "llvm/IR/Module.h"
98 #include "llvm/IR/Operator.h"
99 #include "llvm/IR/ValueHandle.h"
100 #include "llvm/Pass.h"
101 #include "llvm/Support/CommandLine.h"
102 #include "llvm/Support/Debug.h"
103 #include "llvm/Support/ErrorHandling.h"
104 #include "llvm/Support/raw_ostream.h"
105 #include <vector>
106 using namespace llvm;
107
108 #define DEBUG_TYPE "mergefunc"
109
110 STATISTIC(NumFunctionsMerged, "Number of functions merged");
111 STATISTIC(NumThunksWritten, "Number of thunks generated");
112 STATISTIC(NumAliasesWritten, "Number of aliases generated");
113 STATISTIC(NumDoubleWeak, "Number of new functions created");
114
115 static cl::opt<unsigned> NumFunctionsForSanityCheck(
116 "mergefunc-sanity",
117 cl::desc("How many functions in module could be used for "
118 "MergeFunctions pass sanity check. "
119 "'0' disables this check. Works only with '-debug' key."),
120 cl::init(0), cl::Hidden);
121
122 namespace {
123
124 /// FunctionComparator - Compares two functions to determine whether or not
125 /// they will generate machine code with the same behaviour. DataLayout is
126 /// used if available. The comparator always fails conservatively (erring on the
127 /// side of claiming that two functions are different).
128 class FunctionComparator {
129 public:
FunctionComparator(const DataLayout * DL,const Function * F1,const Function * F2)130 FunctionComparator(const DataLayout *DL, const Function *F1,
131 const Function *F2)
132 : FnL(F1), FnR(F2), DL(DL) {}
133
134 /// Test whether the two functions have equivalent behaviour.
135 int compare();
136
137 private:
138 /// Test whether two basic blocks have equivalent behaviour.
139 int compare(const BasicBlock *BBL, const BasicBlock *BBR);
140
141 /// Constants comparison.
142 /// Its analog to lexicographical comparison between hypothetical numbers
143 /// of next format:
144 /// <bitcastability-trait><raw-bit-contents>
145 ///
146 /// 1. Bitcastability.
147 /// Check whether L's type could be losslessly bitcasted to R's type.
148 /// On this stage method, in case when lossless bitcast is not possible
149 /// method returns -1 or 1, thus also defining which type is greater in
150 /// context of bitcastability.
151 /// Stage 0: If types are equal in terms of cmpTypes, then we can go straight
152 /// to the contents comparison.
153 /// If types differ, remember types comparison result and check
154 /// whether we still can bitcast types.
155 /// Stage 1: Types that satisfies isFirstClassType conditions are always
156 /// greater then others.
157 /// Stage 2: Vector is greater then non-vector.
158 /// If both types are vectors, then vector with greater bitwidth is
159 /// greater.
160 /// If both types are vectors with the same bitwidth, then types
161 /// are bitcastable, and we can skip other stages, and go to contents
162 /// comparison.
163 /// Stage 3: Pointer types are greater than non-pointers. If both types are
164 /// pointers of the same address space - go to contents comparison.
165 /// Different address spaces: pointer with greater address space is
166 /// greater.
167 /// Stage 4: Types are neither vectors, nor pointers. And they differ.
168 /// We don't know how to bitcast them. So, we better don't do it,
169 /// and return types comparison result (so it determines the
170 /// relationship among constants we don't know how to bitcast).
171 ///
172 /// Just for clearance, let's see how the set of constants could look
173 /// on single dimension axis:
174 ///
175 /// [NFCT], [FCT, "others"], [FCT, pointers], [FCT, vectors]
176 /// Where: NFCT - Not a FirstClassType
177 /// FCT - FirstClassTyp:
178 ///
179 /// 2. Compare raw contents.
180 /// It ignores types on this stage and only compares bits from L and R.
181 /// Returns 0, if L and R has equivalent contents.
182 /// -1 or 1 if values are different.
183 /// Pretty trivial:
184 /// 2.1. If contents are numbers, compare numbers.
185 /// Ints with greater bitwidth are greater. Ints with same bitwidths
186 /// compared by their contents.
187 /// 2.2. "And so on". Just to avoid discrepancies with comments
188 /// perhaps it would be better to read the implementation itself.
189 /// 3. And again about overall picture. Let's look back at how the ordered set
190 /// of constants will look like:
191 /// [NFCT], [FCT, "others"], [FCT, pointers], [FCT, vectors]
192 ///
193 /// Now look, what could be inside [FCT, "others"], for example:
194 /// [FCT, "others"] =
195 /// [
196 /// [double 0.1], [double 1.23],
197 /// [i32 1], [i32 2],
198 /// { double 1.0 }, ; StructTyID, NumElements = 1
199 /// { i32 1 }, ; StructTyID, NumElements = 1
200 /// { double 1, i32 1 }, ; StructTyID, NumElements = 2
201 /// { i32 1, double 1 } ; StructTyID, NumElements = 2
202 /// ]
203 ///
204 /// Let's explain the order. Float numbers will be less than integers, just
205 /// because of cmpType terms: FloatTyID < IntegerTyID.
206 /// Floats (with same fltSemantics) are sorted according to their value.
207 /// Then you can see integers, and they are, like a floats,
208 /// could be easy sorted among each others.
209 /// The structures. Structures are grouped at the tail, again because of their
210 /// TypeID: StructTyID > IntegerTyID > FloatTyID.
211 /// Structures with greater number of elements are greater. Structures with
212 /// greater elements going first are greater.
213 /// The same logic with vectors, arrays and other possible complex types.
214 ///
215 /// Bitcastable constants.
216 /// Let's assume, that some constant, belongs to some group of
217 /// "so-called-equal" values with different types, and at the same time
218 /// belongs to another group of constants with equal types
219 /// and "really" equal values.
220 ///
221 /// Now, prove that this is impossible:
222 ///
223 /// If constant A with type TyA is bitcastable to B with type TyB, then:
224 /// 1. All constants with equal types to TyA, are bitcastable to B. Since
225 /// those should be vectors (if TyA is vector), pointers
226 /// (if TyA is pointer), or else (if TyA equal to TyB), those types should
227 /// be equal to TyB.
228 /// 2. All constants with non-equal, but bitcastable types to TyA, are
229 /// bitcastable to B.
230 /// Once again, just because we allow it to vectors and pointers only.
231 /// This statement could be expanded as below:
232 /// 2.1. All vectors with equal bitwidth to vector A, has equal bitwidth to
233 /// vector B, and thus bitcastable to B as well.
234 /// 2.2. All pointers of the same address space, no matter what they point to,
235 /// bitcastable. So if C is pointer, it could be bitcasted to A and to B.
236 /// So any constant equal or bitcastable to A is equal or bitcastable to B.
237 /// QED.
238 ///
239 /// In another words, for pointers and vectors, we ignore top-level type and
240 /// look at their particular properties (bit-width for vectors, and
241 /// address space for pointers).
242 /// If these properties are equal - compare their contents.
243 int cmpConstants(const Constant *L, const Constant *R);
244
245 /// Assign or look up previously assigned numbers for the two values, and
246 /// return whether the numbers are equal. Numbers are assigned in the order
247 /// visited.
248 /// Comparison order:
249 /// Stage 0: Value that is function itself is always greater then others.
250 /// If left and right values are references to their functions, then
251 /// they are equal.
252 /// Stage 1: Constants are greater than non-constants.
253 /// If both left and right are constants, then the result of
254 /// cmpConstants is used as cmpValues result.
255 /// Stage 2: InlineAsm instances are greater than others. If both left and
256 /// right are InlineAsm instances, InlineAsm* pointers casted to
257 /// integers and compared as numbers.
258 /// Stage 3: For all other cases we compare order we meet these values in
259 /// their functions. If right value was met first during scanning,
260 /// then left value is greater.
261 /// In another words, we compare serial numbers, for more details
262 /// see comments for sn_mapL and sn_mapR.
263 int cmpValues(const Value *L, const Value *R);
264
265 /// Compare two Instructions for equivalence, similar to
266 /// Instruction::isSameOperationAs but with modifications to the type
267 /// comparison.
268 /// Stages are listed in "most significant stage first" order:
269 /// On each stage below, we do comparison between some left and right
270 /// operation parts. If parts are non-equal, we assign parts comparison
271 /// result to the operation comparison result and exit from method.
272 /// Otherwise we proceed to the next stage.
273 /// Stages:
274 /// 1. Operations opcodes. Compared as numbers.
275 /// 2. Number of operands.
276 /// 3. Operation types. Compared with cmpType method.
277 /// 4. Compare operation subclass optional data as stream of bytes:
278 /// just convert it to integers and call cmpNumbers.
279 /// 5. Compare in operation operand types with cmpType in
280 /// most significant operand first order.
281 /// 6. Last stage. Check operations for some specific attributes.
282 /// For example, for Load it would be:
283 /// 6.1.Load: volatile (as boolean flag)
284 /// 6.2.Load: alignment (as integer numbers)
285 /// 6.3.Load: synch-scope (as integer numbers)
286 /// 6.4.Load: range metadata (as integer numbers)
287 /// On this stage its better to see the code, since its not more than 10-15
288 /// strings for particular instruction, and could change sometimes.
289 int cmpOperations(const Instruction *L, const Instruction *R) const;
290
291 /// Compare two GEPs for equivalent pointer arithmetic.
292 /// Parts to be compared for each comparison stage,
293 /// most significant stage first:
294 /// 1. Address space. As numbers.
295 /// 2. Constant offset, (if "DataLayout *DL" field is not NULL,
296 /// using GEPOperator::accumulateConstantOffset method).
297 /// 3. Pointer operand type (using cmpType method).
298 /// 4. Number of operands.
299 /// 5. Compare operands, using cmpValues method.
300 int cmpGEPs(const GEPOperator *GEPL, const GEPOperator *GEPR);
cmpGEPs(const GetElementPtrInst * GEPL,const GetElementPtrInst * GEPR)301 int cmpGEPs(const GetElementPtrInst *GEPL, const GetElementPtrInst *GEPR) {
302 return cmpGEPs(cast<GEPOperator>(GEPL), cast<GEPOperator>(GEPR));
303 }
304
305 /// cmpType - compares two types,
306 /// defines total ordering among the types set.
307 ///
308 /// Return values:
309 /// 0 if types are equal,
310 /// -1 if Left is less than Right,
311 /// +1 if Left is greater than Right.
312 ///
313 /// Description:
314 /// Comparison is broken onto stages. Like in lexicographical comparison
315 /// stage coming first has higher priority.
316 /// On each explanation stage keep in mind total ordering properties.
317 ///
318 /// 0. Before comparison we coerce pointer types of 0 address space to
319 /// integer.
320 /// We also don't bother with same type at left and right, so
321 /// just return 0 in this case.
322 ///
323 /// 1. If types are of different kind (different type IDs).
324 /// Return result of type IDs comparison, treating them as numbers.
325 /// 2. If types are vectors or integers, compare Type* values as numbers.
326 /// 3. Types has same ID, so check whether they belongs to the next group:
327 /// * Void
328 /// * Float
329 /// * Double
330 /// * X86_FP80
331 /// * FP128
332 /// * PPC_FP128
333 /// * Label
334 /// * Metadata
335 /// If so - return 0, yes - we can treat these types as equal only because
336 /// their IDs are same.
337 /// 4. If Left and Right are pointers, return result of address space
338 /// comparison (numbers comparison). We can treat pointer types of same
339 /// address space as equal.
340 /// 5. If types are complex.
341 /// Then both Left and Right are to be expanded and their element types will
342 /// be checked with the same way. If we get Res != 0 on some stage, return it.
343 /// Otherwise return 0.
344 /// 6. For all other cases put llvm_unreachable.
345 int cmpTypes(Type *TyL, Type *TyR) const;
346
347 int cmpNumbers(uint64_t L, uint64_t R) const;
348
349 int cmpAPInts(const APInt &L, const APInt &R) const;
350 int cmpAPFloats(const APFloat &L, const APFloat &R) const;
351 int cmpStrings(StringRef L, StringRef R) const;
352 int cmpAttrs(const AttributeSet L, const AttributeSet R) const;
353
354 // The two functions undergoing comparison.
355 const Function *FnL, *FnR;
356
357 const DataLayout *DL;
358
359 /// Assign serial numbers to values from left function, and values from
360 /// right function.
361 /// Explanation:
362 /// Being comparing functions we need to compare values we meet at left and
363 /// right sides.
364 /// Its easy to sort things out for external values. It just should be
365 /// the same value at left and right.
366 /// But for local values (those were introduced inside function body)
367 /// we have to ensure they were introduced at exactly the same place,
368 /// and plays the same role.
369 /// Let's assign serial number to each value when we meet it first time.
370 /// Values that were met at same place will be with same serial numbers.
371 /// In this case it would be good to explain few points about values assigned
372 /// to BBs and other ways of implementation (see below).
373 ///
374 /// 1. Safety of BB reordering.
375 /// It's safe to change the order of BasicBlocks in function.
376 /// Relationship with other functions and serial numbering will not be
377 /// changed in this case.
378 /// As follows from FunctionComparator::compare(), we do CFG walk: we start
379 /// from the entry, and then take each terminator. So it doesn't matter how in
380 /// fact BBs are ordered in function. And since cmpValues are called during
381 /// this walk, the numbering depends only on how BBs located inside the CFG.
382 /// So the answer is - yes. We will get the same numbering.
383 ///
384 /// 2. Impossibility to use dominance properties of values.
385 /// If we compare two instruction operands: first is usage of local
386 /// variable AL from function FL, and second is usage of local variable AR
387 /// from FR, we could compare their origins and check whether they are
388 /// defined at the same place.
389 /// But, we are still not able to compare operands of PHI nodes, since those
390 /// could be operands from further BBs we didn't scan yet.
391 /// So it's impossible to use dominance properties in general.
392 DenseMap<const Value*, int> sn_mapL, sn_mapR;
393 };
394
395 class FunctionNode {
396 AssertingVH<Function> F;
397 const DataLayout *DL;
398
399 public:
FunctionNode(Function * F,const DataLayout * DL)400 FunctionNode(Function *F, const DataLayout *DL) : F(F), DL(DL) {}
getFunc() const401 Function *getFunc() const { return F; }
release()402 void release() { F = 0; }
operator <(const FunctionNode & RHS) const403 bool operator<(const FunctionNode &RHS) const {
404 return (FunctionComparator(DL, F, RHS.getFunc()).compare()) == -1;
405 }
406 };
407 }
408
cmpNumbers(uint64_t L,uint64_t R) const409 int FunctionComparator::cmpNumbers(uint64_t L, uint64_t R) const {
410 if (L < R) return -1;
411 if (L > R) return 1;
412 return 0;
413 }
414
cmpAPInts(const APInt & L,const APInt & R) const415 int FunctionComparator::cmpAPInts(const APInt &L, const APInt &R) const {
416 if (int Res = cmpNumbers(L.getBitWidth(), R.getBitWidth()))
417 return Res;
418 if (L.ugt(R)) return 1;
419 if (R.ugt(L)) return -1;
420 return 0;
421 }
422
cmpAPFloats(const APFloat & L,const APFloat & R) const423 int FunctionComparator::cmpAPFloats(const APFloat &L, const APFloat &R) const {
424 if (int Res = cmpNumbers((uint64_t)&L.getSemantics(),
425 (uint64_t)&R.getSemantics()))
426 return Res;
427 return cmpAPInts(L.bitcastToAPInt(), R.bitcastToAPInt());
428 }
429
cmpStrings(StringRef L,StringRef R) const430 int FunctionComparator::cmpStrings(StringRef L, StringRef R) const {
431 // Prevent heavy comparison, compare sizes first.
432 if (int Res = cmpNumbers(L.size(), R.size()))
433 return Res;
434
435 // Compare strings lexicographically only when it is necessary: only when
436 // strings are equal in size.
437 return L.compare(R);
438 }
439
cmpAttrs(const AttributeSet L,const AttributeSet R) const440 int FunctionComparator::cmpAttrs(const AttributeSet L,
441 const AttributeSet R) const {
442 if (int Res = cmpNumbers(L.getNumSlots(), R.getNumSlots()))
443 return Res;
444
445 for (unsigned i = 0, e = L.getNumSlots(); i != e; ++i) {
446 AttributeSet::iterator LI = L.begin(i), LE = L.end(i), RI = R.begin(i),
447 RE = R.end(i);
448 for (; LI != LE && RI != RE; ++LI, ++RI) {
449 Attribute LA = *LI;
450 Attribute RA = *RI;
451 if (LA < RA)
452 return -1;
453 if (RA < LA)
454 return 1;
455 }
456 if (LI != LE)
457 return 1;
458 if (RI != RE)
459 return -1;
460 }
461 return 0;
462 }
463
464 /// Constants comparison:
465 /// 1. Check whether type of L constant could be losslessly bitcasted to R
466 /// type.
467 /// 2. Compare constant contents.
468 /// For more details see declaration comments.
cmpConstants(const Constant * L,const Constant * R)469 int FunctionComparator::cmpConstants(const Constant *L, const Constant *R) {
470
471 Type *TyL = L->getType();
472 Type *TyR = R->getType();
473
474 // Check whether types are bitcastable. This part is just re-factored
475 // Type::canLosslesslyBitCastTo method, but instead of returning true/false,
476 // we also pack into result which type is "less" for us.
477 int TypesRes = cmpTypes(TyL, TyR);
478 if (TypesRes != 0) {
479 // Types are different, but check whether we can bitcast them.
480 if (!TyL->isFirstClassType()) {
481 if (TyR->isFirstClassType())
482 return -1;
483 // Neither TyL nor TyR are values of first class type. Return the result
484 // of comparing the types
485 return TypesRes;
486 }
487 if (!TyR->isFirstClassType()) {
488 if (TyL->isFirstClassType())
489 return 1;
490 return TypesRes;
491 }
492
493 // Vector -> Vector conversions are always lossless if the two vector types
494 // have the same size, otherwise not.
495 unsigned TyLWidth = 0;
496 unsigned TyRWidth = 0;
497
498 if (const VectorType *VecTyL = dyn_cast<VectorType>(TyL))
499 TyLWidth = VecTyL->getBitWidth();
500 if (const VectorType *VecTyR = dyn_cast<VectorType>(TyR))
501 TyRWidth = VecTyR->getBitWidth();
502
503 if (TyLWidth != TyRWidth)
504 return cmpNumbers(TyLWidth, TyRWidth);
505
506 // Zero bit-width means neither TyL nor TyR are vectors.
507 if (!TyLWidth) {
508 PointerType *PTyL = dyn_cast<PointerType>(TyL);
509 PointerType *PTyR = dyn_cast<PointerType>(TyR);
510 if (PTyL && PTyR) {
511 unsigned AddrSpaceL = PTyL->getAddressSpace();
512 unsigned AddrSpaceR = PTyR->getAddressSpace();
513 if (int Res = cmpNumbers(AddrSpaceL, AddrSpaceR))
514 return Res;
515 }
516 if (PTyL)
517 return 1;
518 if (PTyR)
519 return -1;
520
521 // TyL and TyR aren't vectors, nor pointers. We don't know how to
522 // bitcast them.
523 return TypesRes;
524 }
525 }
526
527 // OK, types are bitcastable, now check constant contents.
528
529 if (L->isNullValue() && R->isNullValue())
530 return TypesRes;
531 if (L->isNullValue() && !R->isNullValue())
532 return 1;
533 if (!L->isNullValue() && R->isNullValue())
534 return -1;
535
536 if (int Res = cmpNumbers(L->getValueID(), R->getValueID()))
537 return Res;
538
539 switch (L->getValueID()) {
540 case Value::UndefValueVal: return TypesRes;
541 case Value::ConstantIntVal: {
542 const APInt &LInt = cast<ConstantInt>(L)->getValue();
543 const APInt &RInt = cast<ConstantInt>(R)->getValue();
544 return cmpAPInts(LInt, RInt);
545 }
546 case Value::ConstantFPVal: {
547 const APFloat &LAPF = cast<ConstantFP>(L)->getValueAPF();
548 const APFloat &RAPF = cast<ConstantFP>(R)->getValueAPF();
549 return cmpAPFloats(LAPF, RAPF);
550 }
551 case Value::ConstantArrayVal: {
552 const ConstantArray *LA = cast<ConstantArray>(L);
553 const ConstantArray *RA = cast<ConstantArray>(R);
554 uint64_t NumElementsL = cast<ArrayType>(TyL)->getNumElements();
555 uint64_t NumElementsR = cast<ArrayType>(TyR)->getNumElements();
556 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
557 return Res;
558 for (uint64_t i = 0; i < NumElementsL; ++i) {
559 if (int Res = cmpConstants(cast<Constant>(LA->getOperand(i)),
560 cast<Constant>(RA->getOperand(i))))
561 return Res;
562 }
563 return 0;
564 }
565 case Value::ConstantStructVal: {
566 const ConstantStruct *LS = cast<ConstantStruct>(L);
567 const ConstantStruct *RS = cast<ConstantStruct>(R);
568 unsigned NumElementsL = cast<StructType>(TyL)->getNumElements();
569 unsigned NumElementsR = cast<StructType>(TyR)->getNumElements();
570 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
571 return Res;
572 for (unsigned i = 0; i != NumElementsL; ++i) {
573 if (int Res = cmpConstants(cast<Constant>(LS->getOperand(i)),
574 cast<Constant>(RS->getOperand(i))))
575 return Res;
576 }
577 return 0;
578 }
579 case Value::ConstantVectorVal: {
580 const ConstantVector *LV = cast<ConstantVector>(L);
581 const ConstantVector *RV = cast<ConstantVector>(R);
582 unsigned NumElementsL = cast<VectorType>(TyL)->getNumElements();
583 unsigned NumElementsR = cast<VectorType>(TyR)->getNumElements();
584 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
585 return Res;
586 for (uint64_t i = 0; i < NumElementsL; ++i) {
587 if (int Res = cmpConstants(cast<Constant>(LV->getOperand(i)),
588 cast<Constant>(RV->getOperand(i))))
589 return Res;
590 }
591 return 0;
592 }
593 case Value::ConstantExprVal: {
594 const ConstantExpr *LE = cast<ConstantExpr>(L);
595 const ConstantExpr *RE = cast<ConstantExpr>(R);
596 unsigned NumOperandsL = LE->getNumOperands();
597 unsigned NumOperandsR = RE->getNumOperands();
598 if (int Res = cmpNumbers(NumOperandsL, NumOperandsR))
599 return Res;
600 for (unsigned i = 0; i < NumOperandsL; ++i) {
601 if (int Res = cmpConstants(cast<Constant>(LE->getOperand(i)),
602 cast<Constant>(RE->getOperand(i))))
603 return Res;
604 }
605 return 0;
606 }
607 case Value::FunctionVal:
608 case Value::GlobalVariableVal:
609 case Value::GlobalAliasVal:
610 default: // Unknown constant, cast L and R pointers to numbers and compare.
611 return cmpNumbers((uint64_t)L, (uint64_t)R);
612 }
613 }
614
615 /// cmpType - compares two types,
616 /// defines total ordering among the types set.
617 /// See method declaration comments for more details.
cmpTypes(Type * TyL,Type * TyR) const618 int FunctionComparator::cmpTypes(Type *TyL, Type *TyR) const {
619
620 PointerType *PTyL = dyn_cast<PointerType>(TyL);
621 PointerType *PTyR = dyn_cast<PointerType>(TyR);
622
623 if (DL) {
624 if (PTyL && PTyL->getAddressSpace() == 0) TyL = DL->getIntPtrType(TyL);
625 if (PTyR && PTyR->getAddressSpace() == 0) TyR = DL->getIntPtrType(TyR);
626 }
627
628 if (TyL == TyR)
629 return 0;
630
631 if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
632 return Res;
633
634 switch (TyL->getTypeID()) {
635 default:
636 llvm_unreachable("Unknown type!");
637 // Fall through in Release mode.
638 case Type::IntegerTyID:
639 case Type::VectorTyID:
640 // TyL == TyR would have returned true earlier.
641 return cmpNumbers((uint64_t)TyL, (uint64_t)TyR);
642
643 case Type::VoidTyID:
644 case Type::FloatTyID:
645 case Type::DoubleTyID:
646 case Type::X86_FP80TyID:
647 case Type::FP128TyID:
648 case Type::PPC_FP128TyID:
649 case Type::LabelTyID:
650 case Type::MetadataTyID:
651 return 0;
652
653 case Type::PointerTyID: {
654 assert(PTyL && PTyR && "Both types must be pointers here.");
655 return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
656 }
657
658 case Type::StructTyID: {
659 StructType *STyL = cast<StructType>(TyL);
660 StructType *STyR = cast<StructType>(TyR);
661 if (STyL->getNumElements() != STyR->getNumElements())
662 return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
663
664 if (STyL->isPacked() != STyR->isPacked())
665 return cmpNumbers(STyL->isPacked(), STyR->isPacked());
666
667 for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
668 if (int Res = cmpTypes(STyL->getElementType(i), STyR->getElementType(i)))
669 return Res;
670 }
671 return 0;
672 }
673
674 case Type::FunctionTyID: {
675 FunctionType *FTyL = cast<FunctionType>(TyL);
676 FunctionType *FTyR = cast<FunctionType>(TyR);
677 if (FTyL->getNumParams() != FTyR->getNumParams())
678 return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
679
680 if (FTyL->isVarArg() != FTyR->isVarArg())
681 return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
682
683 if (int Res = cmpTypes(FTyL->getReturnType(), FTyR->getReturnType()))
684 return Res;
685
686 for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
687 if (int Res = cmpTypes(FTyL->getParamType(i), FTyR->getParamType(i)))
688 return Res;
689 }
690 return 0;
691 }
692
693 case Type::ArrayTyID: {
694 ArrayType *ATyL = cast<ArrayType>(TyL);
695 ArrayType *ATyR = cast<ArrayType>(TyR);
696 if (ATyL->getNumElements() != ATyR->getNumElements())
697 return cmpNumbers(ATyL->getNumElements(), ATyR->getNumElements());
698 return cmpTypes(ATyL->getElementType(), ATyR->getElementType());
699 }
700 }
701 }
702
703 // Determine whether the two operations are the same except that pointer-to-A
704 // and pointer-to-B are equivalent. This should be kept in sync with
705 // Instruction::isSameOperationAs.
706 // Read method declaration comments for more details.
cmpOperations(const Instruction * L,const Instruction * R) const707 int FunctionComparator::cmpOperations(const Instruction *L,
708 const Instruction *R) const {
709 // Differences from Instruction::isSameOperationAs:
710 // * replace type comparison with calls to isEquivalentType.
711 // * we test for I->hasSameSubclassOptionalData (nuw/nsw/tail) at the top
712 // * because of the above, we don't test for the tail bit on calls later on
713 if (int Res = cmpNumbers(L->getOpcode(), R->getOpcode()))
714 return Res;
715
716 if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
717 return Res;
718
719 if (int Res = cmpTypes(L->getType(), R->getType()))
720 return Res;
721
722 if (int Res = cmpNumbers(L->getRawSubclassOptionalData(),
723 R->getRawSubclassOptionalData()))
724 return Res;
725
726 // We have two instructions of identical opcode and #operands. Check to see
727 // if all operands are the same type
728 for (unsigned i = 0, e = L->getNumOperands(); i != e; ++i) {
729 if (int Res =
730 cmpTypes(L->getOperand(i)->getType(), R->getOperand(i)->getType()))
731 return Res;
732 }
733
734 // Check special state that is a part of some instructions.
735 if (const LoadInst *LI = dyn_cast<LoadInst>(L)) {
736 if (int Res = cmpNumbers(LI->isVolatile(), cast<LoadInst>(R)->isVolatile()))
737 return Res;
738 if (int Res =
739 cmpNumbers(LI->getAlignment(), cast<LoadInst>(R)->getAlignment()))
740 return Res;
741 if (int Res =
742 cmpNumbers(LI->getOrdering(), cast<LoadInst>(R)->getOrdering()))
743 return Res;
744 if (int Res =
745 cmpNumbers(LI->getSynchScope(), cast<LoadInst>(R)->getSynchScope()))
746 return Res;
747 return cmpNumbers((uint64_t)LI->getMetadata(LLVMContext::MD_range),
748 (uint64_t)cast<LoadInst>(R)->getMetadata(LLVMContext::MD_range));
749 }
750 if (const StoreInst *SI = dyn_cast<StoreInst>(L)) {
751 if (int Res =
752 cmpNumbers(SI->isVolatile(), cast<StoreInst>(R)->isVolatile()))
753 return Res;
754 if (int Res =
755 cmpNumbers(SI->getAlignment(), cast<StoreInst>(R)->getAlignment()))
756 return Res;
757 if (int Res =
758 cmpNumbers(SI->getOrdering(), cast<StoreInst>(R)->getOrdering()))
759 return Res;
760 return cmpNumbers(SI->getSynchScope(), cast<StoreInst>(R)->getSynchScope());
761 }
762 if (const CmpInst *CI = dyn_cast<CmpInst>(L))
763 return cmpNumbers(CI->getPredicate(), cast<CmpInst>(R)->getPredicate());
764 if (const CallInst *CI = dyn_cast<CallInst>(L)) {
765 if (int Res = cmpNumbers(CI->getCallingConv(),
766 cast<CallInst>(R)->getCallingConv()))
767 return Res;
768 if (int Res =
769 cmpAttrs(CI->getAttributes(), cast<CallInst>(R)->getAttributes()))
770 return Res;
771 return cmpNumbers(
772 (uint64_t)CI->getMetadata(LLVMContext::MD_range),
773 (uint64_t)cast<CallInst>(R)->getMetadata(LLVMContext::MD_range));
774 }
775 if (const InvokeInst *CI = dyn_cast<InvokeInst>(L)) {
776 if (int Res = cmpNumbers(CI->getCallingConv(),
777 cast<InvokeInst>(R)->getCallingConv()))
778 return Res;
779 if (int Res =
780 cmpAttrs(CI->getAttributes(), cast<InvokeInst>(R)->getAttributes()))
781 return Res;
782 return cmpNumbers(
783 (uint64_t)CI->getMetadata(LLVMContext::MD_range),
784 (uint64_t)cast<InvokeInst>(R)->getMetadata(LLVMContext::MD_range));
785 }
786 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(L)) {
787 ArrayRef<unsigned> LIndices = IVI->getIndices();
788 ArrayRef<unsigned> RIndices = cast<InsertValueInst>(R)->getIndices();
789 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
790 return Res;
791 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
792 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
793 return Res;
794 }
795 }
796 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(L)) {
797 ArrayRef<unsigned> LIndices = EVI->getIndices();
798 ArrayRef<unsigned> RIndices = cast<ExtractValueInst>(R)->getIndices();
799 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
800 return Res;
801 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
802 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
803 return Res;
804 }
805 }
806 if (const FenceInst *FI = dyn_cast<FenceInst>(L)) {
807 if (int Res =
808 cmpNumbers(FI->getOrdering(), cast<FenceInst>(R)->getOrdering()))
809 return Res;
810 return cmpNumbers(FI->getSynchScope(), cast<FenceInst>(R)->getSynchScope());
811 }
812
813 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(L)) {
814 if (int Res = cmpNumbers(CXI->isVolatile(),
815 cast<AtomicCmpXchgInst>(R)->isVolatile()))
816 return Res;
817 if (int Res = cmpNumbers(CXI->isWeak(),
818 cast<AtomicCmpXchgInst>(R)->isWeak()))
819 return Res;
820 if (int Res = cmpNumbers(CXI->getSuccessOrdering(),
821 cast<AtomicCmpXchgInst>(R)->getSuccessOrdering()))
822 return Res;
823 if (int Res = cmpNumbers(CXI->getFailureOrdering(),
824 cast<AtomicCmpXchgInst>(R)->getFailureOrdering()))
825 return Res;
826 return cmpNumbers(CXI->getSynchScope(),
827 cast<AtomicCmpXchgInst>(R)->getSynchScope());
828 }
829 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(L)) {
830 if (int Res = cmpNumbers(RMWI->getOperation(),
831 cast<AtomicRMWInst>(R)->getOperation()))
832 return Res;
833 if (int Res = cmpNumbers(RMWI->isVolatile(),
834 cast<AtomicRMWInst>(R)->isVolatile()))
835 return Res;
836 if (int Res = cmpNumbers(RMWI->getOrdering(),
837 cast<AtomicRMWInst>(R)->getOrdering()))
838 return Res;
839 return cmpNumbers(RMWI->getSynchScope(),
840 cast<AtomicRMWInst>(R)->getSynchScope());
841 }
842 return 0;
843 }
844
845 // Determine whether two GEP operations perform the same underlying arithmetic.
846 // Read method declaration comments for more details.
cmpGEPs(const GEPOperator * GEPL,const GEPOperator * GEPR)847 int FunctionComparator::cmpGEPs(const GEPOperator *GEPL,
848 const GEPOperator *GEPR) {
849
850 unsigned int ASL = GEPL->getPointerAddressSpace();
851 unsigned int ASR = GEPR->getPointerAddressSpace();
852
853 if (int Res = cmpNumbers(ASL, ASR))
854 return Res;
855
856 // When we have target data, we can reduce the GEP down to the value in bytes
857 // added to the address.
858 if (DL) {
859 unsigned BitWidth = DL->getPointerSizeInBits(ASL);
860 APInt OffsetL(BitWidth, 0), OffsetR(BitWidth, 0);
861 if (GEPL->accumulateConstantOffset(*DL, OffsetL) &&
862 GEPR->accumulateConstantOffset(*DL, OffsetR))
863 return cmpAPInts(OffsetL, OffsetR);
864 }
865
866 if (int Res = cmpNumbers((uint64_t)GEPL->getPointerOperand()->getType(),
867 (uint64_t)GEPR->getPointerOperand()->getType()))
868 return Res;
869
870 if (int Res = cmpNumbers(GEPL->getNumOperands(), GEPR->getNumOperands()))
871 return Res;
872
873 for (unsigned i = 0, e = GEPL->getNumOperands(); i != e; ++i) {
874 if (int Res = cmpValues(GEPL->getOperand(i), GEPR->getOperand(i)))
875 return Res;
876 }
877
878 return 0;
879 }
880
881 /// Compare two values used by the two functions under pair-wise comparison. If
882 /// this is the first time the values are seen, they're added to the mapping so
883 /// that we will detect mismatches on next use.
884 /// See comments in declaration for more details.
cmpValues(const Value * L,const Value * R)885 int FunctionComparator::cmpValues(const Value *L, const Value *R) {
886 // Catch self-reference case.
887 if (L == FnL) {
888 if (R == FnR)
889 return 0;
890 return -1;
891 }
892 if (R == FnR) {
893 if (L == FnL)
894 return 0;
895 return 1;
896 }
897
898 const Constant *ConstL = dyn_cast<Constant>(L);
899 const Constant *ConstR = dyn_cast<Constant>(R);
900 if (ConstL && ConstR) {
901 if (L == R)
902 return 0;
903 return cmpConstants(ConstL, ConstR);
904 }
905
906 if (ConstL)
907 return 1;
908 if (ConstR)
909 return -1;
910
911 const InlineAsm *InlineAsmL = dyn_cast<InlineAsm>(L);
912 const InlineAsm *InlineAsmR = dyn_cast<InlineAsm>(R);
913
914 if (InlineAsmL && InlineAsmR)
915 return cmpNumbers((uint64_t)L, (uint64_t)R);
916 if (InlineAsmL)
917 return 1;
918 if (InlineAsmR)
919 return -1;
920
921 auto LeftSN = sn_mapL.insert(std::make_pair(L, sn_mapL.size())),
922 RightSN = sn_mapR.insert(std::make_pair(R, sn_mapR.size()));
923
924 return cmpNumbers(LeftSN.first->second, RightSN.first->second);
925 }
926 // Test whether two basic blocks have equivalent behaviour.
compare(const BasicBlock * BBL,const BasicBlock * BBR)927 int FunctionComparator::compare(const BasicBlock *BBL, const BasicBlock *BBR) {
928 BasicBlock::const_iterator InstL = BBL->begin(), InstLE = BBL->end();
929 BasicBlock::const_iterator InstR = BBR->begin(), InstRE = BBR->end();
930
931 do {
932 if (int Res = cmpValues(InstL, InstR))
933 return Res;
934
935 const GetElementPtrInst *GEPL = dyn_cast<GetElementPtrInst>(InstL);
936 const GetElementPtrInst *GEPR = dyn_cast<GetElementPtrInst>(InstR);
937
938 if (GEPL && !GEPR)
939 return 1;
940 if (GEPR && !GEPL)
941 return -1;
942
943 if (GEPL && GEPR) {
944 if (int Res =
945 cmpValues(GEPL->getPointerOperand(), GEPR->getPointerOperand()))
946 return Res;
947 if (int Res = cmpGEPs(GEPL, GEPR))
948 return Res;
949 } else {
950 if (int Res = cmpOperations(InstL, InstR))
951 return Res;
952 assert(InstL->getNumOperands() == InstR->getNumOperands());
953
954 for (unsigned i = 0, e = InstL->getNumOperands(); i != e; ++i) {
955 Value *OpL = InstL->getOperand(i);
956 Value *OpR = InstR->getOperand(i);
957 if (int Res = cmpValues(OpL, OpR))
958 return Res;
959 if (int Res = cmpNumbers(OpL->getValueID(), OpR->getValueID()))
960 return Res;
961 // TODO: Already checked in cmpOperation
962 if (int Res = cmpTypes(OpL->getType(), OpR->getType()))
963 return Res;
964 }
965 }
966
967 ++InstL, ++InstR;
968 } while (InstL != InstLE && InstR != InstRE);
969
970 if (InstL != InstLE && InstR == InstRE)
971 return 1;
972 if (InstL == InstLE && InstR != InstRE)
973 return -1;
974 return 0;
975 }
976
977 // Test whether the two functions have equivalent behaviour.
compare()978 int FunctionComparator::compare() {
979
980 sn_mapL.clear();
981 sn_mapR.clear();
982
983 if (int Res = cmpAttrs(FnL->getAttributes(), FnR->getAttributes()))
984 return Res;
985
986 if (int Res = cmpNumbers(FnL->hasGC(), FnR->hasGC()))
987 return Res;
988
989 if (FnL->hasGC()) {
990 if (int Res = cmpNumbers((uint64_t)FnL->getGC(), (uint64_t)FnR->getGC()))
991 return Res;
992 }
993
994 if (int Res = cmpNumbers(FnL->hasSection(), FnR->hasSection()))
995 return Res;
996
997 if (FnL->hasSection()) {
998 if (int Res = cmpStrings(FnL->getSection(), FnR->getSection()))
999 return Res;
1000 }
1001
1002 if (int Res = cmpNumbers(FnL->isVarArg(), FnR->isVarArg()))
1003 return Res;
1004
1005 // TODO: if it's internal and only used in direct calls, we could handle this
1006 // case too.
1007 if (int Res = cmpNumbers(FnL->getCallingConv(), FnR->getCallingConv()))
1008 return Res;
1009
1010 if (int Res = cmpTypes(FnL->getFunctionType(), FnR->getFunctionType()))
1011 return Res;
1012
1013 assert(FnL->arg_size() == FnR->arg_size() &&
1014 "Identically typed functions have different numbers of args!");
1015
1016 // Visit the arguments so that they get enumerated in the order they're
1017 // passed in.
1018 for (Function::const_arg_iterator ArgLI = FnL->arg_begin(),
1019 ArgRI = FnR->arg_begin(),
1020 ArgLE = FnL->arg_end();
1021 ArgLI != ArgLE; ++ArgLI, ++ArgRI) {
1022 if (cmpValues(ArgLI, ArgRI) != 0)
1023 llvm_unreachable("Arguments repeat!");
1024 }
1025
1026 // We do a CFG-ordered walk since the actual ordering of the blocks in the
1027 // linked list is immaterial. Our walk starts at the entry block for both
1028 // functions, then takes each block from each terminator in order. As an
1029 // artifact, this also means that unreachable blocks are ignored.
1030 SmallVector<const BasicBlock *, 8> FnLBBs, FnRBBs;
1031 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
1032
1033 FnLBBs.push_back(&FnL->getEntryBlock());
1034 FnRBBs.push_back(&FnR->getEntryBlock());
1035
1036 VisitedBBs.insert(FnLBBs[0]);
1037 while (!FnLBBs.empty()) {
1038 const BasicBlock *BBL = FnLBBs.pop_back_val();
1039 const BasicBlock *BBR = FnRBBs.pop_back_val();
1040
1041 if (int Res = cmpValues(BBL, BBR))
1042 return Res;
1043
1044 if (int Res = compare(BBL, BBR))
1045 return Res;
1046
1047 const TerminatorInst *TermL = BBL->getTerminator();
1048 const TerminatorInst *TermR = BBR->getTerminator();
1049
1050 assert(TermL->getNumSuccessors() == TermR->getNumSuccessors());
1051 for (unsigned i = 0, e = TermL->getNumSuccessors(); i != e; ++i) {
1052 if (!VisitedBBs.insert(TermL->getSuccessor(i)).second)
1053 continue;
1054
1055 FnLBBs.push_back(TermL->getSuccessor(i));
1056 FnRBBs.push_back(TermR->getSuccessor(i));
1057 }
1058 }
1059 return 0;
1060 }
1061
1062 namespace {
1063
1064 /// MergeFunctions finds functions which will generate identical machine code,
1065 /// by considering all pointer types to be equivalent. Once identified,
1066 /// MergeFunctions will fold them by replacing a call to one to a call to a
1067 /// bitcast of the other.
1068 ///
1069 class MergeFunctions : public ModulePass {
1070 public:
1071 static char ID;
MergeFunctions()1072 MergeFunctions()
1073 : ModulePass(ID), HasGlobalAliases(false) {
1074 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
1075 }
1076
1077 bool runOnModule(Module &M) override;
1078
1079 private:
1080 typedef std::set<FunctionNode> FnTreeType;
1081
1082 /// A work queue of functions that may have been modified and should be
1083 /// analyzed again.
1084 std::vector<WeakVH> Deferred;
1085
1086 /// Checks the rules of order relation introduced among functions set.
1087 /// Returns true, if sanity check has been passed, and false if failed.
1088 bool doSanityCheck(std::vector<WeakVH> &Worklist);
1089
1090 /// Insert a ComparableFunction into the FnTree, or merge it away if it's
1091 /// equal to one that's already present.
1092 bool insert(Function *NewFunction);
1093
1094 /// Remove a Function from the FnTree and queue it up for a second sweep of
1095 /// analysis.
1096 void remove(Function *F);
1097
1098 /// Find the functions that use this Value and remove them from FnTree and
1099 /// queue the functions.
1100 void removeUsers(Value *V);
1101
1102 /// Replace all direct calls of Old with calls of New. Will bitcast New if
1103 /// necessary to make types match.
1104 void replaceDirectCallers(Function *Old, Function *New);
1105
1106 /// Merge two equivalent functions. Upon completion, G may be deleted, or may
1107 /// be converted into a thunk. In either case, it should never be visited
1108 /// again.
1109 void mergeTwoFunctions(Function *F, Function *G);
1110
1111 /// Replace G with a thunk or an alias to F. Deletes G.
1112 void writeThunkOrAlias(Function *F, Function *G);
1113
1114 /// Replace G with a simple tail call to bitcast(F). Also replace direct uses
1115 /// of G with bitcast(F). Deletes G.
1116 void writeThunk(Function *F, Function *G);
1117
1118 /// Replace G with an alias to F. Deletes G.
1119 void writeAlias(Function *F, Function *G);
1120
1121 /// The set of all distinct functions. Use the insert() and remove() methods
1122 /// to modify it.
1123 FnTreeType FnTree;
1124
1125 /// DataLayout for more accurate GEP comparisons. May be NULL.
1126 const DataLayout *DL;
1127
1128 /// Whether or not the target supports global aliases.
1129 bool HasGlobalAliases;
1130 };
1131
1132 } // end anonymous namespace
1133
1134 char MergeFunctions::ID = 0;
1135 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
1136
createMergeFunctionsPass()1137 ModulePass *llvm::createMergeFunctionsPass() {
1138 return new MergeFunctions();
1139 }
1140
doSanityCheck(std::vector<WeakVH> & Worklist)1141 bool MergeFunctions::doSanityCheck(std::vector<WeakVH> &Worklist) {
1142 if (const unsigned Max = NumFunctionsForSanityCheck) {
1143 unsigned TripleNumber = 0;
1144 bool Valid = true;
1145
1146 dbgs() << "MERGEFUNC-SANITY: Started for first " << Max << " functions.\n";
1147
1148 unsigned i = 0;
1149 for (std::vector<WeakVH>::iterator I = Worklist.begin(), E = Worklist.end();
1150 I != E && i < Max; ++I, ++i) {
1151 unsigned j = i;
1152 for (std::vector<WeakVH>::iterator J = I; J != E && j < Max; ++J, ++j) {
1153 Function *F1 = cast<Function>(*I);
1154 Function *F2 = cast<Function>(*J);
1155 int Res1 = FunctionComparator(DL, F1, F2).compare();
1156 int Res2 = FunctionComparator(DL, F2, F1).compare();
1157
1158 // If F1 <= F2, then F2 >= F1, otherwise report failure.
1159 if (Res1 != -Res2) {
1160 dbgs() << "MERGEFUNC-SANITY: Non-symmetric; triple: " << TripleNumber
1161 << "\n";
1162 F1->dump();
1163 F2->dump();
1164 Valid = false;
1165 }
1166
1167 if (Res1 == 0)
1168 continue;
1169
1170 unsigned k = j;
1171 for (std::vector<WeakVH>::iterator K = J; K != E && k < Max;
1172 ++k, ++K, ++TripleNumber) {
1173 if (K == J)
1174 continue;
1175
1176 Function *F3 = cast<Function>(*K);
1177 int Res3 = FunctionComparator(DL, F1, F3).compare();
1178 int Res4 = FunctionComparator(DL, F2, F3).compare();
1179
1180 bool Transitive = true;
1181
1182 if (Res1 != 0 && Res1 == Res4) {
1183 // F1 > F2, F2 > F3 => F1 > F3
1184 Transitive = Res3 == Res1;
1185 } else if (Res3 != 0 && Res3 == -Res4) {
1186 // F1 > F3, F3 > F2 => F1 > F2
1187 Transitive = Res3 == Res1;
1188 } else if (Res4 != 0 && -Res3 == Res4) {
1189 // F2 > F3, F3 > F1 => F2 > F1
1190 Transitive = Res4 == -Res1;
1191 }
1192
1193 if (!Transitive) {
1194 dbgs() << "MERGEFUNC-SANITY: Non-transitive; triple: "
1195 << TripleNumber << "\n";
1196 dbgs() << "Res1, Res3, Res4: " << Res1 << ", " << Res3 << ", "
1197 << Res4 << "\n";
1198 F1->dump();
1199 F2->dump();
1200 F3->dump();
1201 Valid = false;
1202 }
1203 }
1204 }
1205 }
1206
1207 dbgs() << "MERGEFUNC-SANITY: " << (Valid ? "Passed." : "Failed.") << "\n";
1208 return Valid;
1209 }
1210 return true;
1211 }
1212
runOnModule(Module & M)1213 bool MergeFunctions::runOnModule(Module &M) {
1214 bool Changed = false;
1215 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1216 DL = DLP ? &DLP->getDataLayout() : nullptr;
1217
1218 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
1219 if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage())
1220 Deferred.push_back(WeakVH(I));
1221 }
1222
1223 do {
1224 std::vector<WeakVH> Worklist;
1225 Deferred.swap(Worklist);
1226
1227 DEBUG(doSanityCheck(Worklist));
1228
1229 DEBUG(dbgs() << "size of module: " << M.size() << '\n');
1230 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
1231
1232 // Insert only strong functions and merge them. Strong function merging
1233 // always deletes one of them.
1234 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
1235 E = Worklist.end(); I != E; ++I) {
1236 if (!*I) continue;
1237 Function *F = cast<Function>(*I);
1238 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
1239 !F->mayBeOverridden()) {
1240 Changed |= insert(F);
1241 }
1242 }
1243
1244 // Insert only weak functions and merge them. By doing these second we
1245 // create thunks to the strong function when possible. When two weak
1246 // functions are identical, we create a new strong function with two weak
1247 // weak thunks to it which are identical but not mergable.
1248 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
1249 E = Worklist.end(); I != E; ++I) {
1250 if (!*I) continue;
1251 Function *F = cast<Function>(*I);
1252 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
1253 F->mayBeOverridden()) {
1254 Changed |= insert(F);
1255 }
1256 }
1257 DEBUG(dbgs() << "size of FnTree: " << FnTree.size() << '\n');
1258 } while (!Deferred.empty());
1259
1260 FnTree.clear();
1261
1262 return Changed;
1263 }
1264
1265 // Replace direct callers of Old with New.
replaceDirectCallers(Function * Old,Function * New)1266 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
1267 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
1268 for (auto UI = Old->use_begin(), UE = Old->use_end(); UI != UE;) {
1269 Use *U = &*UI;
1270 ++UI;
1271 CallSite CS(U->getUser());
1272 if (CS && CS.isCallee(U)) {
1273 remove(CS.getInstruction()->getParent()->getParent());
1274 U->set(BitcastNew);
1275 }
1276 }
1277 }
1278
1279 // Replace G with an alias to F if possible, or else a thunk to F. Deletes G.
writeThunkOrAlias(Function * F,Function * G)1280 void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) {
1281 if (HasGlobalAliases && G->hasUnnamedAddr()) {
1282 if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
1283 G->hasWeakLinkage()) {
1284 writeAlias(F, G);
1285 return;
1286 }
1287 }
1288
1289 writeThunk(F, G);
1290 }
1291
1292 // Helper for writeThunk,
1293 // Selects proper bitcast operation,
1294 // but a bit simpler then CastInst::getCastOpcode.
createCast(IRBuilder<false> & Builder,Value * V,Type * DestTy)1295 static Value *createCast(IRBuilder<false> &Builder, Value *V, Type *DestTy) {
1296 Type *SrcTy = V->getType();
1297 if (SrcTy->isStructTy()) {
1298 assert(DestTy->isStructTy());
1299 assert(SrcTy->getStructNumElements() == DestTy->getStructNumElements());
1300 Value *Result = UndefValue::get(DestTy);
1301 for (unsigned int I = 0, E = SrcTy->getStructNumElements(); I < E; ++I) {
1302 Value *Element = createCast(
1303 Builder, Builder.CreateExtractValue(V, makeArrayRef(I)),
1304 DestTy->getStructElementType(I));
1305
1306 Result =
1307 Builder.CreateInsertValue(Result, Element, makeArrayRef(I));
1308 }
1309 return Result;
1310 }
1311 assert(!DestTy->isStructTy());
1312 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
1313 return Builder.CreateIntToPtr(V, DestTy);
1314 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
1315 return Builder.CreatePtrToInt(V, DestTy);
1316 else
1317 return Builder.CreateBitCast(V, DestTy);
1318 }
1319
1320 // Replace G with a simple tail call to bitcast(F). Also replace direct uses
1321 // of G with bitcast(F). Deletes G.
writeThunk(Function * F,Function * G)1322 void MergeFunctions::writeThunk(Function *F, Function *G) {
1323 if (!G->mayBeOverridden()) {
1324 // Redirect direct callers of G to F.
1325 replaceDirectCallers(G, F);
1326 }
1327
1328 // If G was internal then we may have replaced all uses of G with F. If so,
1329 // stop here and delete G. There's no need for a thunk.
1330 if (G->hasLocalLinkage() && G->use_empty()) {
1331 G->eraseFromParent();
1332 return;
1333 }
1334
1335 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
1336 G->getParent());
1337 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
1338 IRBuilder<false> Builder(BB);
1339
1340 SmallVector<Value *, 16> Args;
1341 unsigned i = 0;
1342 FunctionType *FFTy = F->getFunctionType();
1343 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
1344 AI != AE; ++AI) {
1345 Args.push_back(createCast(Builder, (Value*)AI, FFTy->getParamType(i)));
1346 ++i;
1347 }
1348
1349 CallInst *CI = Builder.CreateCall(F, Args);
1350 CI->setTailCall();
1351 CI->setCallingConv(F->getCallingConv());
1352 if (NewG->getReturnType()->isVoidTy()) {
1353 Builder.CreateRetVoid();
1354 } else {
1355 Builder.CreateRet(createCast(Builder, CI, NewG->getReturnType()));
1356 }
1357
1358 NewG->copyAttributesFrom(G);
1359 NewG->takeName(G);
1360 removeUsers(G);
1361 G->replaceAllUsesWith(NewG);
1362 G->eraseFromParent();
1363
1364 DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n');
1365 ++NumThunksWritten;
1366 }
1367
1368 // Replace G with an alias to F and delete G.
writeAlias(Function * F,Function * G)1369 void MergeFunctions::writeAlias(Function *F, Function *G) {
1370 PointerType *PTy = G->getType();
1371 auto *GA = GlobalAlias::create(PTy->getElementType(), PTy->getAddressSpace(),
1372 G->getLinkage(), "", F);
1373 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
1374 GA->takeName(G);
1375 GA->setVisibility(G->getVisibility());
1376 removeUsers(G);
1377 G->replaceAllUsesWith(GA);
1378 G->eraseFromParent();
1379
1380 DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n');
1381 ++NumAliasesWritten;
1382 }
1383
1384 // Merge two equivalent functions. Upon completion, Function G is deleted.
mergeTwoFunctions(Function * F,Function * G)1385 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) {
1386 if (F->mayBeOverridden()) {
1387 assert(G->mayBeOverridden());
1388
1389 if (HasGlobalAliases) {
1390 // Make them both thunks to the same internal function.
1391 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
1392 F->getParent());
1393 H->copyAttributesFrom(F);
1394 H->takeName(F);
1395 removeUsers(F);
1396 F->replaceAllUsesWith(H);
1397
1398 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
1399
1400 writeAlias(F, G);
1401 writeAlias(F, H);
1402
1403 F->setAlignment(MaxAlignment);
1404 F->setLinkage(GlobalValue::PrivateLinkage);
1405 } else {
1406 // We can't merge them. Instead, pick one and update all direct callers
1407 // to call it and hope that we improve the instruction cache hit rate.
1408 replaceDirectCallers(G, F);
1409 }
1410
1411 ++NumDoubleWeak;
1412 } else {
1413 writeThunkOrAlias(F, G);
1414 }
1415
1416 ++NumFunctionsMerged;
1417 }
1418
1419 // Insert a ComparableFunction into the FnTree, or merge it away if equal to one
1420 // that was already inserted.
insert(Function * NewFunction)1421 bool MergeFunctions::insert(Function *NewFunction) {
1422 std::pair<FnTreeType::iterator, bool> Result =
1423 FnTree.insert(FunctionNode(NewFunction, DL));
1424
1425 if (Result.second) {
1426 DEBUG(dbgs() << "Inserting as unique: " << NewFunction->getName() << '\n');
1427 return false;
1428 }
1429
1430 const FunctionNode &OldF = *Result.first;
1431
1432 // Don't merge tiny functions, since it can just end up making the function
1433 // larger.
1434 // FIXME: Should still merge them if they are unnamed_addr and produce an
1435 // alias.
1436 if (NewFunction->size() == 1) {
1437 if (NewFunction->front().size() <= 2) {
1438 DEBUG(dbgs() << NewFunction->getName()
1439 << " is to small to bother merging\n");
1440 return false;
1441 }
1442 }
1443
1444 // Never thunk a strong function to a weak function.
1445 assert(!OldF.getFunc()->mayBeOverridden() || NewFunction->mayBeOverridden());
1446
1447 DEBUG(dbgs() << " " << OldF.getFunc()->getName()
1448 << " == " << NewFunction->getName() << '\n');
1449
1450 Function *DeleteF = NewFunction;
1451 mergeTwoFunctions(OldF.getFunc(), DeleteF);
1452 return true;
1453 }
1454
1455 // Remove a function from FnTree. If it was already in FnTree, add
1456 // it to Deferred so that we'll look at it in the next round.
remove(Function * F)1457 void MergeFunctions::remove(Function *F) {
1458 // We need to make sure we remove F, not a function "equal" to F per the
1459 // function equality comparator.
1460 FnTreeType::iterator found = FnTree.find(FunctionNode(F, DL));
1461 size_t Erased = 0;
1462 if (found != FnTree.end() && found->getFunc() == F) {
1463 Erased = 1;
1464 FnTree.erase(found);
1465 }
1466
1467 if (Erased) {
1468 DEBUG(dbgs() << "Removed " << F->getName()
1469 << " from set and deferred it.\n");
1470 Deferred.push_back(F);
1471 }
1472 }
1473
1474 // For each instruction used by the value, remove() the function that contains
1475 // the instruction. This should happen right before a call to RAUW.
removeUsers(Value * V)1476 void MergeFunctions::removeUsers(Value *V) {
1477 std::vector<Value *> Worklist;
1478 Worklist.push_back(V);
1479 while (!Worklist.empty()) {
1480 Value *V = Worklist.back();
1481 Worklist.pop_back();
1482
1483 for (User *U : V->users()) {
1484 if (Instruction *I = dyn_cast<Instruction>(U)) {
1485 remove(I->getParent()->getParent());
1486 } else if (isa<GlobalValue>(U)) {
1487 // do nothing
1488 } else if (Constant *C = dyn_cast<Constant>(U)) {
1489 for (User *UU : C->users())
1490 Worklist.push_back(UU);
1491 }
1492 }
1493 }
1494 }
1495