1 //===- FunctionComparator.h - Function Comparator -------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the FunctionComparator and GlobalNumberState classes
10 // which are used by the MergeFunctions pass for comparing functions.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Transforms/Utils/FunctionComparator.h"
15 #include "llvm/ADT/APFloat.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/ArrayRef.h"
18 #include "llvm/ADT/Hashing.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/IR/Attributes.h"
22 #include "llvm/IR/BasicBlock.h"
23 #include "llvm/IR/CallSite.h"
24 #include "llvm/IR/Constant.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/GlobalValue.h"
30 #include "llvm/IR/InlineAsm.h"
31 #include "llvm/IR/InstrTypes.h"
32 #include "llvm/IR/Instruction.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/Metadata.h"
36 #include "llvm/IR/Module.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/IR/Type.h"
39 #include "llvm/IR/Value.h"
40 #include "llvm/Support/Casting.h"
41 #include "llvm/Support/Compiler.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include <cassert>
46 #include <cstddef>
47 #include <cstdint>
48 #include <utility>
49 
50 using namespace llvm;
51 
52 #define DEBUG_TYPE "functioncomparator"
53 
54 int FunctionComparator::cmpNumbers(uint64_t L, uint64_t R) const {
55   if (L < R) return -1;
56   if (L > R) return 1;
57   return 0;
58 }
59 
60 int FunctionComparator::cmpOrderings(AtomicOrdering L, AtomicOrdering R) const {
61   if ((int)L < (int)R) return -1;
62   if ((int)L > (int)R) return 1;
63   return 0;
64 }
65 
66 int FunctionComparator::cmpAPInts(const APInt &L, const APInt &R) const {
67   if (int Res = cmpNumbers(L.getBitWidth(), R.getBitWidth()))
68     return Res;
69   if (L.ugt(R)) return 1;
70   if (R.ugt(L)) return -1;
71   return 0;
72 }
73 
74 int FunctionComparator::cmpAPFloats(const APFloat &L, const APFloat &R) const {
75   // Floats are ordered first by semantics (i.e. float, double, half, etc.),
76   // then by value interpreted as a bitstring (aka APInt).
77   const fltSemantics &SL = L.getSemantics(), &SR = R.getSemantics();
78   if (int Res = cmpNumbers(APFloat::semanticsPrecision(SL),
79                            APFloat::semanticsPrecision(SR)))
80     return Res;
81   if (int Res = cmpNumbers(APFloat::semanticsMaxExponent(SL),
82                            APFloat::semanticsMaxExponent(SR)))
83     return Res;
84   if (int Res = cmpNumbers(APFloat::semanticsMinExponent(SL),
85                            APFloat::semanticsMinExponent(SR)))
86     return Res;
87   if (int Res = cmpNumbers(APFloat::semanticsSizeInBits(SL),
88                            APFloat::semanticsSizeInBits(SR)))
89     return Res;
90   return cmpAPInts(L.bitcastToAPInt(), R.bitcastToAPInt());
91 }
92 
93 int FunctionComparator::cmpMem(StringRef L, StringRef R) const {
94   // Prevent heavy comparison, compare sizes first.
95   if (int Res = cmpNumbers(L.size(), R.size()))
96     return Res;
97 
98   // Compare strings lexicographically only when it is necessary: only when
99   // strings are equal in size.
100   return L.compare(R);
101 }
102 
103 int FunctionComparator::cmpAttrs(const AttributeList L,
104                                  const AttributeList R) const {
105   if (int Res = cmpNumbers(L.getNumAttrSets(), R.getNumAttrSets()))
106     return Res;
107 
108   for (unsigned i = L.index_begin(), e = L.index_end(); i != e; ++i) {
109     AttributeSet LAS = L.getAttributes(i);
110     AttributeSet RAS = R.getAttributes(i);
111     AttributeSet::iterator LI = LAS.begin(), LE = LAS.end();
112     AttributeSet::iterator RI = RAS.begin(), RE = RAS.end();
113     for (; LI != LE && RI != RE; ++LI, ++RI) {
114       Attribute LA = *LI;
115       Attribute RA = *RI;
116       if (LA.isTypeAttribute() && RA.isTypeAttribute()) {
117         if (LA.getKindAsEnum() != RA.getKindAsEnum())
118           return cmpNumbers(LA.getKindAsEnum(), RA.getKindAsEnum());
119 
120         Type *TyL = LA.getValueAsType();
121         Type *TyR = RA.getValueAsType();
122         if (TyL && TyR)
123           return cmpTypes(TyL, TyR);
124 
125         // Two pointers, at least one null, so the comparison result is
126         // independent of the value of a real pointer.
127         return cmpNumbers((uint64_t)TyL, (uint64_t)TyR);
128       }
129       if (LA < RA)
130         return -1;
131       if (RA < LA)
132         return 1;
133     }
134     if (LI != LE)
135       return 1;
136     if (RI != RE)
137       return -1;
138   }
139   return 0;
140 }
141 
142 int FunctionComparator::cmpRangeMetadata(const MDNode *L,
143                                          const MDNode *R) const {
144   if (L == R)
145     return 0;
146   if (!L)
147     return -1;
148   if (!R)
149     return 1;
150   // Range metadata is a sequence of numbers. Make sure they are the same
151   // sequence.
152   // TODO: Note that as this is metadata, it is possible to drop and/or merge
153   // this data when considering functions to merge. Thus this comparison would
154   // return 0 (i.e. equivalent), but merging would become more complicated
155   // because the ranges would need to be unioned. It is not likely that
156   // functions differ ONLY in this metadata if they are actually the same
157   // function semantically.
158   if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
159     return Res;
160   for (size_t I = 0; I < L->getNumOperands(); ++I) {
161     ConstantInt *LLow = mdconst::extract<ConstantInt>(L->getOperand(I));
162     ConstantInt *RLow = mdconst::extract<ConstantInt>(R->getOperand(I));
163     if (int Res = cmpAPInts(LLow->getValue(), RLow->getValue()))
164       return Res;
165   }
166   return 0;
167 }
168 
169 int FunctionComparator::cmpOperandBundlesSchema(const Instruction *L,
170                                                 const Instruction *R) const {
171   ImmutableCallSite LCS(L);
172   ImmutableCallSite RCS(R);
173 
174   assert(LCS && RCS && "Must be calls or invokes!");
175   assert(LCS.isCall() == RCS.isCall() && "Can't compare otherwise!");
176 
177   if (int Res =
178           cmpNumbers(LCS.getNumOperandBundles(), RCS.getNumOperandBundles()))
179     return Res;
180 
181   for (unsigned i = 0, e = LCS.getNumOperandBundles(); i != e; ++i) {
182     auto OBL = LCS.getOperandBundleAt(i);
183     auto OBR = RCS.getOperandBundleAt(i);
184 
185     if (int Res = OBL.getTagName().compare(OBR.getTagName()))
186       return Res;
187 
188     if (int Res = cmpNumbers(OBL.Inputs.size(), OBR.Inputs.size()))
189       return Res;
190   }
191 
192   return 0;
193 }
194 
195 /// Constants comparison:
196 /// 1. Check whether type of L constant could be losslessly bitcasted to R
197 /// type.
198 /// 2. Compare constant contents.
199 /// For more details see declaration comments.
200 int FunctionComparator::cmpConstants(const Constant *L,
201                                      const Constant *R) const {
202   Type *TyL = L->getType();
203   Type *TyR = R->getType();
204 
205   // Check whether types are bitcastable. This part is just re-factored
206   // Type::canLosslesslyBitCastTo method, but instead of returning true/false,
207   // we also pack into result which type is "less" for us.
208   int TypesRes = cmpTypes(TyL, TyR);
209   if (TypesRes != 0) {
210     // Types are different, but check whether we can bitcast them.
211     if (!TyL->isFirstClassType()) {
212       if (TyR->isFirstClassType())
213         return -1;
214       // Neither TyL nor TyR are values of first class type. Return the result
215       // of comparing the types
216       return TypesRes;
217     }
218     if (!TyR->isFirstClassType()) {
219       if (TyL->isFirstClassType())
220         return 1;
221       return TypesRes;
222     }
223 
224     // Vector -> Vector conversions are always lossless if the two vector types
225     // have the same size, otherwise not.
226     unsigned TyLWidth = 0;
227     unsigned TyRWidth = 0;
228 
229     if (auto *VecTyL = dyn_cast<VectorType>(TyL))
230       TyLWidth = VecTyL->getBitWidth();
231     if (auto *VecTyR = dyn_cast<VectorType>(TyR))
232       TyRWidth = VecTyR->getBitWidth();
233 
234     if (TyLWidth != TyRWidth)
235       return cmpNumbers(TyLWidth, TyRWidth);
236 
237     // Zero bit-width means neither TyL nor TyR are vectors.
238     if (!TyLWidth) {
239       PointerType *PTyL = dyn_cast<PointerType>(TyL);
240       PointerType *PTyR = dyn_cast<PointerType>(TyR);
241       if (PTyL && PTyR) {
242         unsigned AddrSpaceL = PTyL->getAddressSpace();
243         unsigned AddrSpaceR = PTyR->getAddressSpace();
244         if (int Res = cmpNumbers(AddrSpaceL, AddrSpaceR))
245           return Res;
246       }
247       if (PTyL)
248         return 1;
249       if (PTyR)
250         return -1;
251 
252       // TyL and TyR aren't vectors, nor pointers. We don't know how to
253       // bitcast them.
254       return TypesRes;
255     }
256   }
257 
258   // OK, types are bitcastable, now check constant contents.
259 
260   if (L->isNullValue() && R->isNullValue())
261     return TypesRes;
262   if (L->isNullValue() && !R->isNullValue())
263     return 1;
264   if (!L->isNullValue() && R->isNullValue())
265     return -1;
266 
267   auto GlobalValueL = const_cast<GlobalValue *>(dyn_cast<GlobalValue>(L));
268   auto GlobalValueR = const_cast<GlobalValue *>(dyn_cast<GlobalValue>(R));
269   if (GlobalValueL && GlobalValueR) {
270     return cmpGlobalValues(GlobalValueL, GlobalValueR);
271   }
272 
273   if (int Res = cmpNumbers(L->getValueID(), R->getValueID()))
274     return Res;
275 
276   if (const auto *SeqL = dyn_cast<ConstantDataSequential>(L)) {
277     const auto *SeqR = cast<ConstantDataSequential>(R);
278     // This handles ConstantDataArray and ConstantDataVector. Note that we
279     // compare the two raw data arrays, which might differ depending on the host
280     // endianness. This isn't a problem though, because the endiness of a module
281     // will affect the order of the constants, but this order is the same
282     // for a given input module and host platform.
283     return cmpMem(SeqL->getRawDataValues(), SeqR->getRawDataValues());
284   }
285 
286   switch (L->getValueID()) {
287   case Value::UndefValueVal:
288   case Value::ConstantTokenNoneVal:
289     return TypesRes;
290   case Value::ConstantIntVal: {
291     const APInt &LInt = cast<ConstantInt>(L)->getValue();
292     const APInt &RInt = cast<ConstantInt>(R)->getValue();
293     return cmpAPInts(LInt, RInt);
294   }
295   case Value::ConstantFPVal: {
296     const APFloat &LAPF = cast<ConstantFP>(L)->getValueAPF();
297     const APFloat &RAPF = cast<ConstantFP>(R)->getValueAPF();
298     return cmpAPFloats(LAPF, RAPF);
299   }
300   case Value::ConstantArrayVal: {
301     const ConstantArray *LA = cast<ConstantArray>(L);
302     const ConstantArray *RA = cast<ConstantArray>(R);
303     uint64_t NumElementsL = cast<ArrayType>(TyL)->getNumElements();
304     uint64_t NumElementsR = cast<ArrayType>(TyR)->getNumElements();
305     if (int Res = cmpNumbers(NumElementsL, NumElementsR))
306       return Res;
307     for (uint64_t i = 0; i < NumElementsL; ++i) {
308       if (int Res = cmpConstants(cast<Constant>(LA->getOperand(i)),
309                                  cast<Constant>(RA->getOperand(i))))
310         return Res;
311     }
312     return 0;
313   }
314   case Value::ConstantStructVal: {
315     const ConstantStruct *LS = cast<ConstantStruct>(L);
316     const ConstantStruct *RS = cast<ConstantStruct>(R);
317     unsigned NumElementsL = cast<StructType>(TyL)->getNumElements();
318     unsigned NumElementsR = cast<StructType>(TyR)->getNumElements();
319     if (int Res = cmpNumbers(NumElementsL, NumElementsR))
320       return Res;
321     for (unsigned i = 0; i != NumElementsL; ++i) {
322       if (int Res = cmpConstants(cast<Constant>(LS->getOperand(i)),
323                                  cast<Constant>(RS->getOperand(i))))
324         return Res;
325     }
326     return 0;
327   }
328   case Value::ConstantVectorVal: {
329     const ConstantVector *LV = cast<ConstantVector>(L);
330     const ConstantVector *RV = cast<ConstantVector>(R);
331     unsigned NumElementsL = cast<VectorType>(TyL)->getNumElements();
332     unsigned NumElementsR = cast<VectorType>(TyR)->getNumElements();
333     if (int Res = cmpNumbers(NumElementsL, NumElementsR))
334       return Res;
335     for (uint64_t i = 0; i < NumElementsL; ++i) {
336       if (int Res = cmpConstants(cast<Constant>(LV->getOperand(i)),
337                                  cast<Constant>(RV->getOperand(i))))
338         return Res;
339     }
340     return 0;
341   }
342   case Value::ConstantExprVal: {
343     const ConstantExpr *LE = cast<ConstantExpr>(L);
344     const ConstantExpr *RE = cast<ConstantExpr>(R);
345     unsigned NumOperandsL = LE->getNumOperands();
346     unsigned NumOperandsR = RE->getNumOperands();
347     if (int Res = cmpNumbers(NumOperandsL, NumOperandsR))
348       return Res;
349     for (unsigned i = 0; i < NumOperandsL; ++i) {
350       if (int Res = cmpConstants(cast<Constant>(LE->getOperand(i)),
351                                  cast<Constant>(RE->getOperand(i))))
352         return Res;
353     }
354     return 0;
355   }
356   case Value::BlockAddressVal: {
357     const BlockAddress *LBA = cast<BlockAddress>(L);
358     const BlockAddress *RBA = cast<BlockAddress>(R);
359     if (int Res = cmpValues(LBA->getFunction(), RBA->getFunction()))
360       return Res;
361     if (LBA->getFunction() == RBA->getFunction()) {
362       // They are BBs in the same function. Order by which comes first in the
363       // BB order of the function. This order is deterministic.
364       Function* F = LBA->getFunction();
365       BasicBlock *LBB = LBA->getBasicBlock();
366       BasicBlock *RBB = RBA->getBasicBlock();
367       if (LBB == RBB)
368         return 0;
369       for(BasicBlock &BB : F->getBasicBlockList()) {
370         if (&BB == LBB) {
371           assert(&BB != RBB);
372           return -1;
373         }
374         if (&BB == RBB)
375           return 1;
376       }
377       llvm_unreachable("Basic Block Address does not point to a basic block in "
378                        "its function.");
379       return -1;
380     } else {
381       // cmpValues said the functions are the same. So because they aren't
382       // literally the same pointer, they must respectively be the left and
383       // right functions.
384       assert(LBA->getFunction() == FnL && RBA->getFunction() == FnR);
385       // cmpValues will tell us if these are equivalent BasicBlocks, in the
386       // context of their respective functions.
387       return cmpValues(LBA->getBasicBlock(), RBA->getBasicBlock());
388     }
389   }
390   default: // Unknown constant, abort.
391     LLVM_DEBUG(dbgs() << "Looking at valueID " << L->getValueID() << "\n");
392     llvm_unreachable("Constant ValueID not recognized.");
393     return -1;
394   }
395 }
396 
397 int FunctionComparator::cmpGlobalValues(GlobalValue *L, GlobalValue *R) const {
398   uint64_t LNumber = GlobalNumbers->getNumber(L);
399   uint64_t RNumber = GlobalNumbers->getNumber(R);
400   return cmpNumbers(LNumber, RNumber);
401 }
402 
403 /// cmpType - compares two types,
404 /// defines total ordering among the types set.
405 /// See method declaration comments for more details.
406 int FunctionComparator::cmpTypes(Type *TyL, Type *TyR) const {
407   PointerType *PTyL = dyn_cast<PointerType>(TyL);
408   PointerType *PTyR = dyn_cast<PointerType>(TyR);
409 
410   const DataLayout &DL = FnL->getParent()->getDataLayout();
411   if (PTyL && PTyL->getAddressSpace() == 0)
412     TyL = DL.getIntPtrType(TyL);
413   if (PTyR && PTyR->getAddressSpace() == 0)
414     TyR = DL.getIntPtrType(TyR);
415 
416   if (TyL == TyR)
417     return 0;
418 
419   if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
420     return Res;
421 
422   switch (TyL->getTypeID()) {
423   default:
424     llvm_unreachable("Unknown type!");
425   case Type::IntegerTyID:
426     return cmpNumbers(cast<IntegerType>(TyL)->getBitWidth(),
427                       cast<IntegerType>(TyR)->getBitWidth());
428   // TyL == TyR would have returned true earlier, because types are uniqued.
429   case Type::VoidTyID:
430   case Type::FloatTyID:
431   case Type::DoubleTyID:
432   case Type::X86_FP80TyID:
433   case Type::FP128TyID:
434   case Type::PPC_FP128TyID:
435   case Type::LabelTyID:
436   case Type::MetadataTyID:
437   case Type::TokenTyID:
438     return 0;
439 
440   case Type::PointerTyID:
441     assert(PTyL && PTyR && "Both types must be pointers here.");
442     return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
443 
444   case Type::StructTyID: {
445     StructType *STyL = cast<StructType>(TyL);
446     StructType *STyR = cast<StructType>(TyR);
447     if (STyL->getNumElements() != STyR->getNumElements())
448       return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
449 
450     if (STyL->isPacked() != STyR->isPacked())
451       return cmpNumbers(STyL->isPacked(), STyR->isPacked());
452 
453     for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
454       if (int Res = cmpTypes(STyL->getElementType(i), STyR->getElementType(i)))
455         return Res;
456     }
457     return 0;
458   }
459 
460   case Type::FunctionTyID: {
461     FunctionType *FTyL = cast<FunctionType>(TyL);
462     FunctionType *FTyR = cast<FunctionType>(TyR);
463     if (FTyL->getNumParams() != FTyR->getNumParams())
464       return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
465 
466     if (FTyL->isVarArg() != FTyR->isVarArg())
467       return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
468 
469     if (int Res = cmpTypes(FTyL->getReturnType(), FTyR->getReturnType()))
470       return Res;
471 
472     for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
473       if (int Res = cmpTypes(FTyL->getParamType(i), FTyR->getParamType(i)))
474         return Res;
475     }
476     return 0;
477   }
478 
479   case Type::ArrayTyID:
480   case Type::VectorTyID: {
481     auto *STyL = cast<SequentialType>(TyL);
482     auto *STyR = cast<SequentialType>(TyR);
483     if (STyL->getNumElements() != STyR->getNumElements())
484       return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
485     return cmpTypes(STyL->getElementType(), STyR->getElementType());
486   }
487   }
488 }
489 
490 // Determine whether the two operations are the same except that pointer-to-A
491 // and pointer-to-B are equivalent. This should be kept in sync with
492 // Instruction::isSameOperationAs.
493 // Read method declaration comments for more details.
494 int FunctionComparator::cmpOperations(const Instruction *L,
495                                       const Instruction *R,
496                                       bool &needToCmpOperands) const {
497   needToCmpOperands = true;
498   if (int Res = cmpValues(L, R))
499     return Res;
500 
501   // Differences from Instruction::isSameOperationAs:
502   //  * replace type comparison with calls to cmpTypes.
503   //  * we test for I->getRawSubclassOptionalData (nuw/nsw/tail) at the top.
504   //  * because of the above, we don't test for the tail bit on calls later on.
505   if (int Res = cmpNumbers(L->getOpcode(), R->getOpcode()))
506     return Res;
507 
508   if (const GetElementPtrInst *GEPL = dyn_cast<GetElementPtrInst>(L)) {
509     needToCmpOperands = false;
510     const GetElementPtrInst *GEPR = cast<GetElementPtrInst>(R);
511     if (int Res =
512             cmpValues(GEPL->getPointerOperand(), GEPR->getPointerOperand()))
513       return Res;
514     return cmpGEPs(GEPL, GEPR);
515   }
516 
517   if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
518     return Res;
519 
520   if (int Res = cmpTypes(L->getType(), R->getType()))
521     return Res;
522 
523   if (int Res = cmpNumbers(L->getRawSubclassOptionalData(),
524                            R->getRawSubclassOptionalData()))
525     return Res;
526 
527   // We have two instructions of identical opcode and #operands.  Check to see
528   // if all operands are the same type
529   for (unsigned i = 0, e = L->getNumOperands(); i != e; ++i) {
530     if (int Res =
531             cmpTypes(L->getOperand(i)->getType(), R->getOperand(i)->getType()))
532       return Res;
533   }
534 
535   // Check special state that is a part of some instructions.
536   if (const AllocaInst *AI = dyn_cast<AllocaInst>(L)) {
537     if (int Res = cmpTypes(AI->getAllocatedType(),
538                            cast<AllocaInst>(R)->getAllocatedType()))
539       return Res;
540     return cmpNumbers(AI->getAlignment(), cast<AllocaInst>(R)->getAlignment());
541   }
542   if (const LoadInst *LI = dyn_cast<LoadInst>(L)) {
543     if (int Res = cmpNumbers(LI->isVolatile(), cast<LoadInst>(R)->isVolatile()))
544       return Res;
545     if (int Res =
546             cmpNumbers(LI->getAlignment(), cast<LoadInst>(R)->getAlignment()))
547       return Res;
548     if (int Res =
549             cmpOrderings(LI->getOrdering(), cast<LoadInst>(R)->getOrdering()))
550       return Res;
551     if (int Res = cmpNumbers(LI->getSyncScopeID(),
552                              cast<LoadInst>(R)->getSyncScopeID()))
553       return Res;
554     return cmpRangeMetadata(LI->getMetadata(LLVMContext::MD_range),
555         cast<LoadInst>(R)->getMetadata(LLVMContext::MD_range));
556   }
557   if (const StoreInst *SI = dyn_cast<StoreInst>(L)) {
558     if (int Res =
559             cmpNumbers(SI->isVolatile(), cast<StoreInst>(R)->isVolatile()))
560       return Res;
561     if (int Res =
562             cmpNumbers(SI->getAlignment(), cast<StoreInst>(R)->getAlignment()))
563       return Res;
564     if (int Res =
565             cmpOrderings(SI->getOrdering(), cast<StoreInst>(R)->getOrdering()))
566       return Res;
567     return cmpNumbers(SI->getSyncScopeID(),
568                       cast<StoreInst>(R)->getSyncScopeID());
569   }
570   if (const CmpInst *CI = dyn_cast<CmpInst>(L))
571     return cmpNumbers(CI->getPredicate(), cast<CmpInst>(R)->getPredicate());
572   if (auto CSL = CallSite(const_cast<Instruction *>(L))) {
573     auto CSR = CallSite(const_cast<Instruction *>(R));
574     if (int Res = cmpNumbers(CSL.getCallingConv(), CSR.getCallingConv()))
575       return Res;
576     if (int Res = cmpAttrs(CSL.getAttributes(), CSR.getAttributes()))
577       return Res;
578     if (int Res = cmpOperandBundlesSchema(L, R))
579       return Res;
580     if (const CallInst *CI = dyn_cast<CallInst>(L))
581       if (int Res = cmpNumbers(CI->getTailCallKind(),
582                                cast<CallInst>(R)->getTailCallKind()))
583         return Res;
584     return cmpRangeMetadata(L->getMetadata(LLVMContext::MD_range),
585                             R->getMetadata(LLVMContext::MD_range));
586   }
587   if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(L)) {
588     ArrayRef<unsigned> LIndices = IVI->getIndices();
589     ArrayRef<unsigned> RIndices = cast<InsertValueInst>(R)->getIndices();
590     if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
591       return Res;
592     for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
593       if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
594         return Res;
595     }
596     return 0;
597   }
598   if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(L)) {
599     ArrayRef<unsigned> LIndices = EVI->getIndices();
600     ArrayRef<unsigned> RIndices = cast<ExtractValueInst>(R)->getIndices();
601     if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
602       return Res;
603     for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
604       if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
605         return Res;
606     }
607   }
608   if (const FenceInst *FI = dyn_cast<FenceInst>(L)) {
609     if (int Res =
610             cmpOrderings(FI->getOrdering(), cast<FenceInst>(R)->getOrdering()))
611       return Res;
612     return cmpNumbers(FI->getSyncScopeID(),
613                       cast<FenceInst>(R)->getSyncScopeID());
614   }
615   if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(L)) {
616     if (int Res = cmpNumbers(CXI->isVolatile(),
617                              cast<AtomicCmpXchgInst>(R)->isVolatile()))
618       return Res;
619     if (int Res = cmpNumbers(CXI->isWeak(),
620                              cast<AtomicCmpXchgInst>(R)->isWeak()))
621       return Res;
622     if (int Res =
623             cmpOrderings(CXI->getSuccessOrdering(),
624                          cast<AtomicCmpXchgInst>(R)->getSuccessOrdering()))
625       return Res;
626     if (int Res =
627             cmpOrderings(CXI->getFailureOrdering(),
628                          cast<AtomicCmpXchgInst>(R)->getFailureOrdering()))
629       return Res;
630     return cmpNumbers(CXI->getSyncScopeID(),
631                       cast<AtomicCmpXchgInst>(R)->getSyncScopeID());
632   }
633   if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(L)) {
634     if (int Res = cmpNumbers(RMWI->getOperation(),
635                              cast<AtomicRMWInst>(R)->getOperation()))
636       return Res;
637     if (int Res = cmpNumbers(RMWI->isVolatile(),
638                              cast<AtomicRMWInst>(R)->isVolatile()))
639       return Res;
640     if (int Res = cmpOrderings(RMWI->getOrdering(),
641                              cast<AtomicRMWInst>(R)->getOrdering()))
642       return Res;
643     return cmpNumbers(RMWI->getSyncScopeID(),
644                       cast<AtomicRMWInst>(R)->getSyncScopeID());
645   }
646   if (const PHINode *PNL = dyn_cast<PHINode>(L)) {
647     const PHINode *PNR = cast<PHINode>(R);
648     // Ensure that in addition to the incoming values being identical
649     // (checked by the caller of this function), the incoming blocks
650     // are also identical.
651     for (unsigned i = 0, e = PNL->getNumIncomingValues(); i != e; ++i) {
652       if (int Res =
653               cmpValues(PNL->getIncomingBlock(i), PNR->getIncomingBlock(i)))
654         return Res;
655     }
656   }
657   return 0;
658 }
659 
660 // Determine whether two GEP operations perform the same underlying arithmetic.
661 // Read method declaration comments for more details.
662 int FunctionComparator::cmpGEPs(const GEPOperator *GEPL,
663                                 const GEPOperator *GEPR) const {
664   unsigned int ASL = GEPL->getPointerAddressSpace();
665   unsigned int ASR = GEPR->getPointerAddressSpace();
666 
667   if (int Res = cmpNumbers(ASL, ASR))
668     return Res;
669 
670   // When we have target data, we can reduce the GEP down to the value in bytes
671   // added to the address.
672   const DataLayout &DL = FnL->getParent()->getDataLayout();
673   unsigned BitWidth = DL.getPointerSizeInBits(ASL);
674   APInt OffsetL(BitWidth, 0), OffsetR(BitWidth, 0);
675   if (GEPL->accumulateConstantOffset(DL, OffsetL) &&
676       GEPR->accumulateConstantOffset(DL, OffsetR))
677     return cmpAPInts(OffsetL, OffsetR);
678   if (int Res = cmpTypes(GEPL->getSourceElementType(),
679                          GEPR->getSourceElementType()))
680     return Res;
681 
682   if (int Res = cmpNumbers(GEPL->getNumOperands(), GEPR->getNumOperands()))
683     return Res;
684 
685   for (unsigned i = 0, e = GEPL->getNumOperands(); i != e; ++i) {
686     if (int Res = cmpValues(GEPL->getOperand(i), GEPR->getOperand(i)))
687       return Res;
688   }
689 
690   return 0;
691 }
692 
693 int FunctionComparator::cmpInlineAsm(const InlineAsm *L,
694                                      const InlineAsm *R) const {
695   // InlineAsm's are uniqued. If they are the same pointer, obviously they are
696   // the same, otherwise compare the fields.
697   if (L == R)
698     return 0;
699   if (int Res = cmpTypes(L->getFunctionType(), R->getFunctionType()))
700     return Res;
701   if (int Res = cmpMem(L->getAsmString(), R->getAsmString()))
702     return Res;
703   if (int Res = cmpMem(L->getConstraintString(), R->getConstraintString()))
704     return Res;
705   if (int Res = cmpNumbers(L->hasSideEffects(), R->hasSideEffects()))
706     return Res;
707   if (int Res = cmpNumbers(L->isAlignStack(), R->isAlignStack()))
708     return Res;
709   if (int Res = cmpNumbers(L->getDialect(), R->getDialect()))
710     return Res;
711   assert(L->getFunctionType() != R->getFunctionType());
712   return 0;
713 }
714 
715 /// Compare two values used by the two functions under pair-wise comparison. If
716 /// this is the first time the values are seen, they're added to the mapping so
717 /// that we will detect mismatches on next use.
718 /// See comments in declaration for more details.
719 int FunctionComparator::cmpValues(const Value *L, const Value *R) const {
720   // Catch self-reference case.
721   if (L == FnL) {
722     if (R == FnR)
723       return 0;
724     return -1;
725   }
726   if (R == FnR) {
727     if (L == FnL)
728       return 0;
729     return 1;
730   }
731 
732   const Constant *ConstL = dyn_cast<Constant>(L);
733   const Constant *ConstR = dyn_cast<Constant>(R);
734   if (ConstL && ConstR) {
735     if (L == R)
736       return 0;
737     return cmpConstants(ConstL, ConstR);
738   }
739 
740   if (ConstL)
741     return 1;
742   if (ConstR)
743     return -1;
744 
745   const InlineAsm *InlineAsmL = dyn_cast<InlineAsm>(L);
746   const InlineAsm *InlineAsmR = dyn_cast<InlineAsm>(R);
747 
748   if (InlineAsmL && InlineAsmR)
749     return cmpInlineAsm(InlineAsmL, InlineAsmR);
750   if (InlineAsmL)
751     return 1;
752   if (InlineAsmR)
753     return -1;
754 
755   auto LeftSN = sn_mapL.insert(std::make_pair(L, sn_mapL.size())),
756        RightSN = sn_mapR.insert(std::make_pair(R, sn_mapR.size()));
757 
758   return cmpNumbers(LeftSN.first->second, RightSN.first->second);
759 }
760 
761 // Test whether two basic blocks have equivalent behaviour.
762 int FunctionComparator::cmpBasicBlocks(const BasicBlock *BBL,
763                                        const BasicBlock *BBR) const {
764   BasicBlock::const_iterator InstL = BBL->begin(), InstLE = BBL->end();
765   BasicBlock::const_iterator InstR = BBR->begin(), InstRE = BBR->end();
766 
767   do {
768     bool needToCmpOperands = true;
769     if (int Res = cmpOperations(&*InstL, &*InstR, needToCmpOperands))
770       return Res;
771     if (needToCmpOperands) {
772       assert(InstL->getNumOperands() == InstR->getNumOperands());
773 
774       for (unsigned i = 0, e = InstL->getNumOperands(); i != e; ++i) {
775         Value *OpL = InstL->getOperand(i);
776         Value *OpR = InstR->getOperand(i);
777         if (int Res = cmpValues(OpL, OpR))
778           return Res;
779         // cmpValues should ensure this is true.
780         assert(cmpTypes(OpL->getType(), OpR->getType()) == 0);
781       }
782     }
783 
784     ++InstL;
785     ++InstR;
786   } while (InstL != InstLE && InstR != InstRE);
787 
788   if (InstL != InstLE && InstR == InstRE)
789     return 1;
790   if (InstL == InstLE && InstR != InstRE)
791     return -1;
792   return 0;
793 }
794 
795 int FunctionComparator::compareSignature() const {
796   if (int Res = cmpAttrs(FnL->getAttributes(), FnR->getAttributes()))
797     return Res;
798 
799   if (int Res = cmpNumbers(FnL->hasGC(), FnR->hasGC()))
800     return Res;
801 
802   if (FnL->hasGC()) {
803     if (int Res = cmpMem(FnL->getGC(), FnR->getGC()))
804       return Res;
805   }
806 
807   if (int Res = cmpNumbers(FnL->hasSection(), FnR->hasSection()))
808     return Res;
809 
810   if (FnL->hasSection()) {
811     if (int Res = cmpMem(FnL->getSection(), FnR->getSection()))
812       return Res;
813   }
814 
815   if (int Res = cmpNumbers(FnL->isVarArg(), FnR->isVarArg()))
816     return Res;
817 
818   // TODO: if it's internal and only used in direct calls, we could handle this
819   // case too.
820   if (int Res = cmpNumbers(FnL->getCallingConv(), FnR->getCallingConv()))
821     return Res;
822 
823   if (int Res = cmpTypes(FnL->getFunctionType(), FnR->getFunctionType()))
824     return Res;
825 
826   assert(FnL->arg_size() == FnR->arg_size() &&
827          "Identically typed functions have different numbers of args!");
828 
829   // Visit the arguments so that they get enumerated in the order they're
830   // passed in.
831   for (Function::const_arg_iterator ArgLI = FnL->arg_begin(),
832        ArgRI = FnR->arg_begin(),
833        ArgLE = FnL->arg_end();
834        ArgLI != ArgLE; ++ArgLI, ++ArgRI) {
835     if (cmpValues(&*ArgLI, &*ArgRI) != 0)
836       llvm_unreachable("Arguments repeat!");
837   }
838   return 0;
839 }
840 
841 // Test whether the two functions have equivalent behaviour.
842 int FunctionComparator::compare() {
843   beginCompare();
844 
845   if (int Res = compareSignature())
846     return Res;
847 
848   // We do a CFG-ordered walk since the actual ordering of the blocks in the
849   // linked list is immaterial. Our walk starts at the entry block for both
850   // functions, then takes each block from each terminator in order. As an
851   // artifact, this also means that unreachable blocks are ignored.
852   SmallVector<const BasicBlock *, 8> FnLBBs, FnRBBs;
853   SmallPtrSet<const BasicBlock *, 32> VisitedBBs; // in terms of F1.
854 
855   FnLBBs.push_back(&FnL->getEntryBlock());
856   FnRBBs.push_back(&FnR->getEntryBlock());
857 
858   VisitedBBs.insert(FnLBBs[0]);
859   while (!FnLBBs.empty()) {
860     const BasicBlock *BBL = FnLBBs.pop_back_val();
861     const BasicBlock *BBR = FnRBBs.pop_back_val();
862 
863     if (int Res = cmpValues(BBL, BBR))
864       return Res;
865 
866     if (int Res = cmpBasicBlocks(BBL, BBR))
867       return Res;
868 
869     const Instruction *TermL = BBL->getTerminator();
870     const Instruction *TermR = BBR->getTerminator();
871 
872     assert(TermL->getNumSuccessors() == TermR->getNumSuccessors());
873     for (unsigned i = 0, e = TermL->getNumSuccessors(); i != e; ++i) {
874       if (!VisitedBBs.insert(TermL->getSuccessor(i)).second)
875         continue;
876 
877       FnLBBs.push_back(TermL->getSuccessor(i));
878       FnRBBs.push_back(TermR->getSuccessor(i));
879     }
880   }
881   return 0;
882 }
883 
884 namespace {
885 
886 // Accumulate the hash of a sequence of 64-bit integers. This is similar to a
887 // hash of a sequence of 64bit ints, but the entire input does not need to be
888 // available at once. This interface is necessary for functionHash because it
889 // needs to accumulate the hash as the structure of the function is traversed
890 // without saving these values to an intermediate buffer. This form of hashing
891 // is not often needed, as usually the object to hash is just read from a
892 // buffer.
893 class HashAccumulator64 {
894   uint64_t Hash;
895 
896 public:
897   // Initialize to random constant, so the state isn't zero.
898   HashAccumulator64() { Hash = 0x6acaa36bef8325c5ULL; }
899 
900   void add(uint64_t V) {
901      Hash = hashing::detail::hash_16_bytes(Hash, V);
902   }
903 
904   // No finishing is required, because the entire hash value is used.
905   uint64_t getHash() { return Hash; }
906 };
907 
908 } // end anonymous namespace
909 
910 // A function hash is calculated by considering only the number of arguments and
911 // whether a function is varargs, the order of basic blocks (given by the
912 // successors of each basic block in depth first order), and the order of
913 // opcodes of each instruction within each of these basic blocks. This mirrors
914 // the strategy compare() uses to compare functions by walking the BBs in depth
915 // first order and comparing each instruction in sequence. Because this hash
916 // does not look at the operands, it is insensitive to things such as the
917 // target of calls and the constants used in the function, which makes it useful
918 // when possibly merging functions which are the same modulo constants and call
919 // targets.
920 FunctionComparator::FunctionHash FunctionComparator::functionHash(Function &F) {
921   HashAccumulator64 H;
922   H.add(F.isVarArg());
923   H.add(F.arg_size());
924 
925   SmallVector<const BasicBlock *, 8> BBs;
926   SmallPtrSet<const BasicBlock *, 16> VisitedBBs;
927 
928   // Walk the blocks in the same order as FunctionComparator::cmpBasicBlocks(),
929   // accumulating the hash of the function "structure." (BB and opcode sequence)
930   BBs.push_back(&F.getEntryBlock());
931   VisitedBBs.insert(BBs[0]);
932   while (!BBs.empty()) {
933     const BasicBlock *BB = BBs.pop_back_val();
934     // This random value acts as a block header, as otherwise the partition of
935     // opcodes into BBs wouldn't affect the hash, only the order of the opcodes
936     H.add(45798);
937     for (auto &Inst : *BB) {
938       H.add(Inst.getOpcode());
939     }
940     const Instruction *Term = BB->getTerminator();
941     for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) {
942       if (!VisitedBBs.insert(Term->getSuccessor(i)).second)
943         continue;
944       BBs.push_back(Term->getSuccessor(i));
945     }
946   }
947   return H.getHash();
948 }
949