1 //===- FunctionComparator.h - Function Comparator -------------------------===//
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 file implements the FunctionComparator and GlobalNumberState classes
11 // which are used by the MergeFunctions pass for comparing functions.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "llvm/Transforms/Utils/FunctionComparator.h"
16 #include "llvm/ADT/APFloat.h"
17 #include "llvm/ADT/APInt.h"
18 #include "llvm/ADT/ArrayRef.h"
19 #include "llvm/ADT/Hashing.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/IR/Attributes.h"
23 #include "llvm/IR/BasicBlock.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/Constant.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/GlobalValue.h"
31 #include "llvm/IR/InlineAsm.h"
32 #include "llvm/IR/InstrTypes.h"
33 #include "llvm/IR/Instruction.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/LLVMContext.h"
36 #include "llvm/IR/Metadata.h"
37 #include "llvm/IR/Module.h"
38 #include "llvm/IR/Operator.h"
39 #include "llvm/IR/Type.h"
40 #include "llvm/IR/Value.h"
41 #include "llvm/Support/Casting.h"
42 #include "llvm/Support/Compiler.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include <cassert>
47 #include <cstddef>
48 #include <cstdint>
49 #include <utility>
50
51 using namespace llvm;
52
53 #define DEBUG_TYPE "functioncomparator"
54
cmpNumbers(uint64_t L,uint64_t R) const55 int FunctionComparator::cmpNumbers(uint64_t L, uint64_t R) const {
56 if (L < R) return -1;
57 if (L > R) return 1;
58 return 0;
59 }
60
cmpOrderings(AtomicOrdering L,AtomicOrdering R) const61 int FunctionComparator::cmpOrderings(AtomicOrdering L, AtomicOrdering R) const {
62 if ((int)L < (int)R) return -1;
63 if ((int)L > (int)R) return 1;
64 return 0;
65 }
66
cmpAPInts(const APInt & L,const APInt & R) const67 int FunctionComparator::cmpAPInts(const APInt &L, const APInt &R) const {
68 if (int Res = cmpNumbers(L.getBitWidth(), R.getBitWidth()))
69 return Res;
70 if (L.ugt(R)) return 1;
71 if (R.ugt(L)) return -1;
72 return 0;
73 }
74
cmpAPFloats(const APFloat & L,const APFloat & R) const75 int FunctionComparator::cmpAPFloats(const APFloat &L, const APFloat &R) const {
76 // Floats are ordered first by semantics (i.e. float, double, half, etc.),
77 // then by value interpreted as a bitstring (aka APInt).
78 const fltSemantics &SL = L.getSemantics(), &SR = R.getSemantics();
79 if (int Res = cmpNumbers(APFloat::semanticsPrecision(SL),
80 APFloat::semanticsPrecision(SR)))
81 return Res;
82 if (int Res = cmpNumbers(APFloat::semanticsMaxExponent(SL),
83 APFloat::semanticsMaxExponent(SR)))
84 return Res;
85 if (int Res = cmpNumbers(APFloat::semanticsMinExponent(SL),
86 APFloat::semanticsMinExponent(SR)))
87 return Res;
88 if (int Res = cmpNumbers(APFloat::semanticsSizeInBits(SL),
89 APFloat::semanticsSizeInBits(SR)))
90 return Res;
91 return cmpAPInts(L.bitcastToAPInt(), R.bitcastToAPInt());
92 }
93
cmpMem(StringRef L,StringRef R) const94 int FunctionComparator::cmpMem(StringRef L, StringRef R) const {
95 // Prevent heavy comparison, compare sizes first.
96 if (int Res = cmpNumbers(L.size(), R.size()))
97 return Res;
98
99 // Compare strings lexicographically only when it is necessary: only when
100 // strings are equal in size.
101 return L.compare(R);
102 }
103
cmpAttrs(const AttributeList L,const AttributeList R) const104 int FunctionComparator::cmpAttrs(const AttributeList L,
105 const AttributeList R) const {
106 if (int Res = cmpNumbers(L.getNumAttrSets(), R.getNumAttrSets()))
107 return Res;
108
109 for (unsigned i = L.index_begin(), e = L.index_end(); i != e; ++i) {
110 AttributeSet LAS = L.getAttributes(i);
111 AttributeSet RAS = R.getAttributes(i);
112 AttributeSet::iterator LI = LAS.begin(), LE = LAS.end();
113 AttributeSet::iterator RI = RAS.begin(), RE = RAS.end();
114 for (; LI != LE && RI != RE; ++LI, ++RI) {
115 Attribute LA = *LI;
116 Attribute RA = *RI;
117 if (LA < RA)
118 return -1;
119 if (RA < LA)
120 return 1;
121 }
122 if (LI != LE)
123 return 1;
124 if (RI != RE)
125 return -1;
126 }
127 return 0;
128 }
129
cmpRangeMetadata(const MDNode * L,const MDNode * R) const130 int FunctionComparator::cmpRangeMetadata(const MDNode *L,
131 const MDNode *R) const {
132 if (L == R)
133 return 0;
134 if (!L)
135 return -1;
136 if (!R)
137 return 1;
138 // Range metadata is a sequence of numbers. Make sure they are the same
139 // sequence.
140 // TODO: Note that as this is metadata, it is possible to drop and/or merge
141 // this data when considering functions to merge. Thus this comparison would
142 // return 0 (i.e. equivalent), but merging would become more complicated
143 // because the ranges would need to be unioned. It is not likely that
144 // functions differ ONLY in this metadata if they are actually the same
145 // function semantically.
146 if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
147 return Res;
148 for (size_t I = 0; I < L->getNumOperands(); ++I) {
149 ConstantInt *LLow = mdconst::extract<ConstantInt>(L->getOperand(I));
150 ConstantInt *RLow = mdconst::extract<ConstantInt>(R->getOperand(I));
151 if (int Res = cmpAPInts(LLow->getValue(), RLow->getValue()))
152 return Res;
153 }
154 return 0;
155 }
156
cmpOperandBundlesSchema(const Instruction * L,const Instruction * R) const157 int FunctionComparator::cmpOperandBundlesSchema(const Instruction *L,
158 const Instruction *R) const {
159 ImmutableCallSite LCS(L);
160 ImmutableCallSite RCS(R);
161
162 assert(LCS && RCS && "Must be calls or invokes!");
163 assert(LCS.isCall() == RCS.isCall() && "Can't compare otherwise!");
164
165 if (int Res =
166 cmpNumbers(LCS.getNumOperandBundles(), RCS.getNumOperandBundles()))
167 return Res;
168
169 for (unsigned i = 0, e = LCS.getNumOperandBundles(); i != e; ++i) {
170 auto OBL = LCS.getOperandBundleAt(i);
171 auto OBR = RCS.getOperandBundleAt(i);
172
173 if (int Res = OBL.getTagName().compare(OBR.getTagName()))
174 return Res;
175
176 if (int Res = cmpNumbers(OBL.Inputs.size(), OBR.Inputs.size()))
177 return Res;
178 }
179
180 return 0;
181 }
182
183 /// Constants comparison:
184 /// 1. Check whether type of L constant could be losslessly bitcasted to R
185 /// type.
186 /// 2. Compare constant contents.
187 /// For more details see declaration comments.
cmpConstants(const Constant * L,const Constant * R) const188 int FunctionComparator::cmpConstants(const Constant *L,
189 const Constant *R) const {
190 Type *TyL = L->getType();
191 Type *TyR = R->getType();
192
193 // Check whether types are bitcastable. This part is just re-factored
194 // Type::canLosslesslyBitCastTo method, but instead of returning true/false,
195 // we also pack into result which type is "less" for us.
196 int TypesRes = cmpTypes(TyL, TyR);
197 if (TypesRes != 0) {
198 // Types are different, but check whether we can bitcast them.
199 if (!TyL->isFirstClassType()) {
200 if (TyR->isFirstClassType())
201 return -1;
202 // Neither TyL nor TyR are values of first class type. Return the result
203 // of comparing the types
204 return TypesRes;
205 }
206 if (!TyR->isFirstClassType()) {
207 if (TyL->isFirstClassType())
208 return 1;
209 return TypesRes;
210 }
211
212 // Vector -> Vector conversions are always lossless if the two vector types
213 // have the same size, otherwise not.
214 unsigned TyLWidth = 0;
215 unsigned TyRWidth = 0;
216
217 if (auto *VecTyL = dyn_cast<VectorType>(TyL))
218 TyLWidth = VecTyL->getBitWidth();
219 if (auto *VecTyR = dyn_cast<VectorType>(TyR))
220 TyRWidth = VecTyR->getBitWidth();
221
222 if (TyLWidth != TyRWidth)
223 return cmpNumbers(TyLWidth, TyRWidth);
224
225 // Zero bit-width means neither TyL nor TyR are vectors.
226 if (!TyLWidth) {
227 PointerType *PTyL = dyn_cast<PointerType>(TyL);
228 PointerType *PTyR = dyn_cast<PointerType>(TyR);
229 if (PTyL && PTyR) {
230 unsigned AddrSpaceL = PTyL->getAddressSpace();
231 unsigned AddrSpaceR = PTyR->getAddressSpace();
232 if (int Res = cmpNumbers(AddrSpaceL, AddrSpaceR))
233 return Res;
234 }
235 if (PTyL)
236 return 1;
237 if (PTyR)
238 return -1;
239
240 // TyL and TyR aren't vectors, nor pointers. We don't know how to
241 // bitcast them.
242 return TypesRes;
243 }
244 }
245
246 // OK, types are bitcastable, now check constant contents.
247
248 if (L->isNullValue() && R->isNullValue())
249 return TypesRes;
250 if (L->isNullValue() && !R->isNullValue())
251 return 1;
252 if (!L->isNullValue() && R->isNullValue())
253 return -1;
254
255 auto GlobalValueL = const_cast<GlobalValue *>(dyn_cast<GlobalValue>(L));
256 auto GlobalValueR = const_cast<GlobalValue *>(dyn_cast<GlobalValue>(R));
257 if (GlobalValueL && GlobalValueR) {
258 return cmpGlobalValues(GlobalValueL, GlobalValueR);
259 }
260
261 if (int Res = cmpNumbers(L->getValueID(), R->getValueID()))
262 return Res;
263
264 if (const auto *SeqL = dyn_cast<ConstantDataSequential>(L)) {
265 const auto *SeqR = cast<ConstantDataSequential>(R);
266 // This handles ConstantDataArray and ConstantDataVector. Note that we
267 // compare the two raw data arrays, which might differ depending on the host
268 // endianness. This isn't a problem though, because the endiness of a module
269 // will affect the order of the constants, but this order is the same
270 // for a given input module and host platform.
271 return cmpMem(SeqL->getRawDataValues(), SeqR->getRawDataValues());
272 }
273
274 switch (L->getValueID()) {
275 case Value::UndefValueVal:
276 case Value::ConstantTokenNoneVal:
277 return TypesRes;
278 case Value::ConstantIntVal: {
279 const APInt &LInt = cast<ConstantInt>(L)->getValue();
280 const APInt &RInt = cast<ConstantInt>(R)->getValue();
281 return cmpAPInts(LInt, RInt);
282 }
283 case Value::ConstantFPVal: {
284 const APFloat &LAPF = cast<ConstantFP>(L)->getValueAPF();
285 const APFloat &RAPF = cast<ConstantFP>(R)->getValueAPF();
286 return cmpAPFloats(LAPF, RAPF);
287 }
288 case Value::ConstantArrayVal: {
289 const ConstantArray *LA = cast<ConstantArray>(L);
290 const ConstantArray *RA = cast<ConstantArray>(R);
291 uint64_t NumElementsL = cast<ArrayType>(TyL)->getNumElements();
292 uint64_t NumElementsR = cast<ArrayType>(TyR)->getNumElements();
293 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
294 return Res;
295 for (uint64_t i = 0; i < NumElementsL; ++i) {
296 if (int Res = cmpConstants(cast<Constant>(LA->getOperand(i)),
297 cast<Constant>(RA->getOperand(i))))
298 return Res;
299 }
300 return 0;
301 }
302 case Value::ConstantStructVal: {
303 const ConstantStruct *LS = cast<ConstantStruct>(L);
304 const ConstantStruct *RS = cast<ConstantStruct>(R);
305 unsigned NumElementsL = cast<StructType>(TyL)->getNumElements();
306 unsigned NumElementsR = cast<StructType>(TyR)->getNumElements();
307 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
308 return Res;
309 for (unsigned i = 0; i != NumElementsL; ++i) {
310 if (int Res = cmpConstants(cast<Constant>(LS->getOperand(i)),
311 cast<Constant>(RS->getOperand(i))))
312 return Res;
313 }
314 return 0;
315 }
316 case Value::ConstantVectorVal: {
317 const ConstantVector *LV = cast<ConstantVector>(L);
318 const ConstantVector *RV = cast<ConstantVector>(R);
319 unsigned NumElementsL = cast<VectorType>(TyL)->getNumElements();
320 unsigned NumElementsR = cast<VectorType>(TyR)->getNumElements();
321 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
322 return Res;
323 for (uint64_t i = 0; i < NumElementsL; ++i) {
324 if (int Res = cmpConstants(cast<Constant>(LV->getOperand(i)),
325 cast<Constant>(RV->getOperand(i))))
326 return Res;
327 }
328 return 0;
329 }
330 case Value::ConstantExprVal: {
331 const ConstantExpr *LE = cast<ConstantExpr>(L);
332 const ConstantExpr *RE = cast<ConstantExpr>(R);
333 unsigned NumOperandsL = LE->getNumOperands();
334 unsigned NumOperandsR = RE->getNumOperands();
335 if (int Res = cmpNumbers(NumOperandsL, NumOperandsR))
336 return Res;
337 for (unsigned i = 0; i < NumOperandsL; ++i) {
338 if (int Res = cmpConstants(cast<Constant>(LE->getOperand(i)),
339 cast<Constant>(RE->getOperand(i))))
340 return Res;
341 }
342 return 0;
343 }
344 case Value::BlockAddressVal: {
345 const BlockAddress *LBA = cast<BlockAddress>(L);
346 const BlockAddress *RBA = cast<BlockAddress>(R);
347 if (int Res = cmpValues(LBA->getFunction(), RBA->getFunction()))
348 return Res;
349 if (LBA->getFunction() == RBA->getFunction()) {
350 // They are BBs in the same function. Order by which comes first in the
351 // BB order of the function. This order is deterministic.
352 Function* F = LBA->getFunction();
353 BasicBlock *LBB = LBA->getBasicBlock();
354 BasicBlock *RBB = RBA->getBasicBlock();
355 if (LBB == RBB)
356 return 0;
357 for(BasicBlock &BB : F->getBasicBlockList()) {
358 if (&BB == LBB) {
359 assert(&BB != RBB);
360 return -1;
361 }
362 if (&BB == RBB)
363 return 1;
364 }
365 llvm_unreachable("Basic Block Address does not point to a basic block in "
366 "its function.");
367 return -1;
368 } else {
369 // cmpValues said the functions are the same. So because they aren't
370 // literally the same pointer, they must respectively be the left and
371 // right functions.
372 assert(LBA->getFunction() == FnL && RBA->getFunction() == FnR);
373 // cmpValues will tell us if these are equivalent BasicBlocks, in the
374 // context of their respective functions.
375 return cmpValues(LBA->getBasicBlock(), RBA->getBasicBlock());
376 }
377 }
378 default: // Unknown constant, abort.
379 LLVM_DEBUG(dbgs() << "Looking at valueID " << L->getValueID() << "\n");
380 llvm_unreachable("Constant ValueID not recognized.");
381 return -1;
382 }
383 }
384
cmpGlobalValues(GlobalValue * L,GlobalValue * R) const385 int FunctionComparator::cmpGlobalValues(GlobalValue *L, GlobalValue *R) const {
386 uint64_t LNumber = GlobalNumbers->getNumber(L);
387 uint64_t RNumber = GlobalNumbers->getNumber(R);
388 return cmpNumbers(LNumber, RNumber);
389 }
390
391 /// cmpType - compares two types,
392 /// defines total ordering among the types set.
393 /// See method declaration comments for more details.
cmpTypes(Type * TyL,Type * TyR) const394 int FunctionComparator::cmpTypes(Type *TyL, Type *TyR) const {
395 PointerType *PTyL = dyn_cast<PointerType>(TyL);
396 PointerType *PTyR = dyn_cast<PointerType>(TyR);
397
398 const DataLayout &DL = FnL->getParent()->getDataLayout();
399 if (PTyL && PTyL->getAddressSpace() == 0)
400 TyL = DL.getIntPtrType(TyL);
401 if (PTyR && PTyR->getAddressSpace() == 0)
402 TyR = DL.getIntPtrType(TyR);
403
404 if (TyL == TyR)
405 return 0;
406
407 if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
408 return Res;
409
410 switch (TyL->getTypeID()) {
411 default:
412 llvm_unreachable("Unknown type!");
413 // Fall through in Release mode.
414 LLVM_FALLTHROUGH;
415 case Type::IntegerTyID:
416 return cmpNumbers(cast<IntegerType>(TyL)->getBitWidth(),
417 cast<IntegerType>(TyR)->getBitWidth());
418 // TyL == TyR would have returned true earlier, because types are uniqued.
419 case Type::VoidTyID:
420 case Type::FloatTyID:
421 case Type::DoubleTyID:
422 case Type::X86_FP80TyID:
423 case Type::FP128TyID:
424 case Type::PPC_FP128TyID:
425 case Type::LabelTyID:
426 case Type::MetadataTyID:
427 case Type::TokenTyID:
428 return 0;
429
430 case Type::PointerTyID:
431 assert(PTyL && PTyR && "Both types must be pointers here.");
432 return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
433
434 case Type::StructTyID: {
435 StructType *STyL = cast<StructType>(TyL);
436 StructType *STyR = cast<StructType>(TyR);
437 if (STyL->getNumElements() != STyR->getNumElements())
438 return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
439
440 if (STyL->isPacked() != STyR->isPacked())
441 return cmpNumbers(STyL->isPacked(), STyR->isPacked());
442
443 for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
444 if (int Res = cmpTypes(STyL->getElementType(i), STyR->getElementType(i)))
445 return Res;
446 }
447 return 0;
448 }
449
450 case Type::FunctionTyID: {
451 FunctionType *FTyL = cast<FunctionType>(TyL);
452 FunctionType *FTyR = cast<FunctionType>(TyR);
453 if (FTyL->getNumParams() != FTyR->getNumParams())
454 return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
455
456 if (FTyL->isVarArg() != FTyR->isVarArg())
457 return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
458
459 if (int Res = cmpTypes(FTyL->getReturnType(), FTyR->getReturnType()))
460 return Res;
461
462 for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
463 if (int Res = cmpTypes(FTyL->getParamType(i), FTyR->getParamType(i)))
464 return Res;
465 }
466 return 0;
467 }
468
469 case Type::ArrayTyID:
470 case Type::VectorTyID: {
471 auto *STyL = cast<SequentialType>(TyL);
472 auto *STyR = cast<SequentialType>(TyR);
473 if (STyL->getNumElements() != STyR->getNumElements())
474 return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
475 return cmpTypes(STyL->getElementType(), STyR->getElementType());
476 }
477 }
478 }
479
480 // Determine whether the two operations are the same except that pointer-to-A
481 // and pointer-to-B are equivalent. This should be kept in sync with
482 // Instruction::isSameOperationAs.
483 // Read method declaration comments for more details.
cmpOperations(const Instruction * L,const Instruction * R,bool & needToCmpOperands) const484 int FunctionComparator::cmpOperations(const Instruction *L,
485 const Instruction *R,
486 bool &needToCmpOperands) const {
487 needToCmpOperands = true;
488 if (int Res = cmpValues(L, R))
489 return Res;
490
491 // Differences from Instruction::isSameOperationAs:
492 // * replace type comparison with calls to cmpTypes.
493 // * we test for I->getRawSubclassOptionalData (nuw/nsw/tail) at the top.
494 // * because of the above, we don't test for the tail bit on calls later on.
495 if (int Res = cmpNumbers(L->getOpcode(), R->getOpcode()))
496 return Res;
497
498 if (const GetElementPtrInst *GEPL = dyn_cast<GetElementPtrInst>(L)) {
499 needToCmpOperands = false;
500 const GetElementPtrInst *GEPR = cast<GetElementPtrInst>(R);
501 if (int Res =
502 cmpValues(GEPL->getPointerOperand(), GEPR->getPointerOperand()))
503 return Res;
504 return cmpGEPs(GEPL, GEPR);
505 }
506
507 if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
508 return Res;
509
510 if (int Res = cmpTypes(L->getType(), R->getType()))
511 return Res;
512
513 if (int Res = cmpNumbers(L->getRawSubclassOptionalData(),
514 R->getRawSubclassOptionalData()))
515 return Res;
516
517 // We have two instructions of identical opcode and #operands. Check to see
518 // if all operands are the same type
519 for (unsigned i = 0, e = L->getNumOperands(); i != e; ++i) {
520 if (int Res =
521 cmpTypes(L->getOperand(i)->getType(), R->getOperand(i)->getType()))
522 return Res;
523 }
524
525 // Check special state that is a part of some instructions.
526 if (const AllocaInst *AI = dyn_cast<AllocaInst>(L)) {
527 if (int Res = cmpTypes(AI->getAllocatedType(),
528 cast<AllocaInst>(R)->getAllocatedType()))
529 return Res;
530 return cmpNumbers(AI->getAlignment(), cast<AllocaInst>(R)->getAlignment());
531 }
532 if (const LoadInst *LI = dyn_cast<LoadInst>(L)) {
533 if (int Res = cmpNumbers(LI->isVolatile(), cast<LoadInst>(R)->isVolatile()))
534 return Res;
535 if (int Res =
536 cmpNumbers(LI->getAlignment(), cast<LoadInst>(R)->getAlignment()))
537 return Res;
538 if (int Res =
539 cmpOrderings(LI->getOrdering(), cast<LoadInst>(R)->getOrdering()))
540 return Res;
541 if (int Res = cmpNumbers(LI->getSyncScopeID(),
542 cast<LoadInst>(R)->getSyncScopeID()))
543 return Res;
544 return cmpRangeMetadata(LI->getMetadata(LLVMContext::MD_range),
545 cast<LoadInst>(R)->getMetadata(LLVMContext::MD_range));
546 }
547 if (const StoreInst *SI = dyn_cast<StoreInst>(L)) {
548 if (int Res =
549 cmpNumbers(SI->isVolatile(), cast<StoreInst>(R)->isVolatile()))
550 return Res;
551 if (int Res =
552 cmpNumbers(SI->getAlignment(), cast<StoreInst>(R)->getAlignment()))
553 return Res;
554 if (int Res =
555 cmpOrderings(SI->getOrdering(), cast<StoreInst>(R)->getOrdering()))
556 return Res;
557 return cmpNumbers(SI->getSyncScopeID(),
558 cast<StoreInst>(R)->getSyncScopeID());
559 }
560 if (const CmpInst *CI = dyn_cast<CmpInst>(L))
561 return cmpNumbers(CI->getPredicate(), cast<CmpInst>(R)->getPredicate());
562 if (const CallInst *CI = dyn_cast<CallInst>(L)) {
563 if (int Res = cmpNumbers(CI->getCallingConv(),
564 cast<CallInst>(R)->getCallingConv()))
565 return Res;
566 if (int Res =
567 cmpAttrs(CI->getAttributes(), cast<CallInst>(R)->getAttributes()))
568 return Res;
569 if (int Res = cmpOperandBundlesSchema(CI, R))
570 return Res;
571 return cmpRangeMetadata(
572 CI->getMetadata(LLVMContext::MD_range),
573 cast<CallInst>(R)->getMetadata(LLVMContext::MD_range));
574 }
575 if (const InvokeInst *II = dyn_cast<InvokeInst>(L)) {
576 if (int Res = cmpNumbers(II->getCallingConv(),
577 cast<InvokeInst>(R)->getCallingConv()))
578 return Res;
579 if (int Res =
580 cmpAttrs(II->getAttributes(), cast<InvokeInst>(R)->getAttributes()))
581 return Res;
582 if (int Res = cmpOperandBundlesSchema(II, R))
583 return Res;
584 return cmpRangeMetadata(
585 II->getMetadata(LLVMContext::MD_range),
586 cast<InvokeInst>(R)->getMetadata(LLVMContext::MD_range));
587 }
588 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(L)) {
589 ArrayRef<unsigned> LIndices = IVI->getIndices();
590 ArrayRef<unsigned> RIndices = cast<InsertValueInst>(R)->getIndices();
591 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
592 return Res;
593 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
594 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
595 return Res;
596 }
597 return 0;
598 }
599 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(L)) {
600 ArrayRef<unsigned> LIndices = EVI->getIndices();
601 ArrayRef<unsigned> RIndices = cast<ExtractValueInst>(R)->getIndices();
602 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
603 return Res;
604 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
605 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
606 return Res;
607 }
608 }
609 if (const FenceInst *FI = dyn_cast<FenceInst>(L)) {
610 if (int Res =
611 cmpOrderings(FI->getOrdering(), cast<FenceInst>(R)->getOrdering()))
612 return Res;
613 return cmpNumbers(FI->getSyncScopeID(),
614 cast<FenceInst>(R)->getSyncScopeID());
615 }
616 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(L)) {
617 if (int Res = cmpNumbers(CXI->isVolatile(),
618 cast<AtomicCmpXchgInst>(R)->isVolatile()))
619 return Res;
620 if (int Res = cmpNumbers(CXI->isWeak(),
621 cast<AtomicCmpXchgInst>(R)->isWeak()))
622 return Res;
623 if (int Res =
624 cmpOrderings(CXI->getSuccessOrdering(),
625 cast<AtomicCmpXchgInst>(R)->getSuccessOrdering()))
626 return Res;
627 if (int Res =
628 cmpOrderings(CXI->getFailureOrdering(),
629 cast<AtomicCmpXchgInst>(R)->getFailureOrdering()))
630 return Res;
631 return cmpNumbers(CXI->getSyncScopeID(),
632 cast<AtomicCmpXchgInst>(R)->getSyncScopeID());
633 }
634 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(L)) {
635 if (int Res = cmpNumbers(RMWI->getOperation(),
636 cast<AtomicRMWInst>(R)->getOperation()))
637 return Res;
638 if (int Res = cmpNumbers(RMWI->isVolatile(),
639 cast<AtomicRMWInst>(R)->isVolatile()))
640 return Res;
641 if (int Res = cmpOrderings(RMWI->getOrdering(),
642 cast<AtomicRMWInst>(R)->getOrdering()))
643 return Res;
644 return cmpNumbers(RMWI->getSyncScopeID(),
645 cast<AtomicRMWInst>(R)->getSyncScopeID());
646 }
647 if (const PHINode *PNL = dyn_cast<PHINode>(L)) {
648 const PHINode *PNR = cast<PHINode>(R);
649 // Ensure that in addition to the incoming values being identical
650 // (checked by the caller of this function), the incoming blocks
651 // are also identical.
652 for (unsigned i = 0, e = PNL->getNumIncomingValues(); i != e; ++i) {
653 if (int Res =
654 cmpValues(PNL->getIncomingBlock(i), PNR->getIncomingBlock(i)))
655 return Res;
656 }
657 }
658 return 0;
659 }
660
661 // Determine whether two GEP operations perform the same underlying arithmetic.
662 // Read method declaration comments for more details.
cmpGEPs(const GEPOperator * GEPL,const GEPOperator * GEPR) const663 int FunctionComparator::cmpGEPs(const GEPOperator *GEPL,
664 const GEPOperator *GEPR) const {
665 unsigned int ASL = GEPL->getPointerAddressSpace();
666 unsigned int ASR = GEPR->getPointerAddressSpace();
667
668 if (int Res = cmpNumbers(ASL, ASR))
669 return Res;
670
671 // When we have target data, we can reduce the GEP down to the value in bytes
672 // added to the address.
673 const DataLayout &DL = FnL->getParent()->getDataLayout();
674 unsigned BitWidth = DL.getPointerSizeInBits(ASL);
675 APInt OffsetL(BitWidth, 0), OffsetR(BitWidth, 0);
676 if (GEPL->accumulateConstantOffset(DL, OffsetL) &&
677 GEPR->accumulateConstantOffset(DL, OffsetR))
678 return cmpAPInts(OffsetL, OffsetR);
679 if (int Res = cmpTypes(GEPL->getSourceElementType(),
680 GEPR->getSourceElementType()))
681 return Res;
682
683 if (int Res = cmpNumbers(GEPL->getNumOperands(), GEPR->getNumOperands()))
684 return Res;
685
686 for (unsigned i = 0, e = GEPL->getNumOperands(); i != e; ++i) {
687 if (int Res = cmpValues(GEPL->getOperand(i), GEPR->getOperand(i)))
688 return Res;
689 }
690
691 return 0;
692 }
693
cmpInlineAsm(const InlineAsm * L,const InlineAsm * R) const694 int FunctionComparator::cmpInlineAsm(const InlineAsm *L,
695 const InlineAsm *R) const {
696 // InlineAsm's are uniqued. If they are the same pointer, obviously they are
697 // the same, otherwise compare the fields.
698 if (L == R)
699 return 0;
700 if (int Res = cmpTypes(L->getFunctionType(), R->getFunctionType()))
701 return Res;
702 if (int Res = cmpMem(L->getAsmString(), R->getAsmString()))
703 return Res;
704 if (int Res = cmpMem(L->getConstraintString(), R->getConstraintString()))
705 return Res;
706 if (int Res = cmpNumbers(L->hasSideEffects(), R->hasSideEffects()))
707 return Res;
708 if (int Res = cmpNumbers(L->isAlignStack(), R->isAlignStack()))
709 return Res;
710 if (int Res = cmpNumbers(L->getDialect(), R->getDialect()))
711 return Res;
712 assert(L->getFunctionType() != R->getFunctionType());
713 return 0;
714 }
715
716 /// Compare two values used by the two functions under pair-wise comparison. If
717 /// this is the first time the values are seen, they're added to the mapping so
718 /// that we will detect mismatches on next use.
719 /// See comments in declaration for more details.
cmpValues(const Value * L,const Value * R) const720 int FunctionComparator::cmpValues(const Value *L, const Value *R) const {
721 // Catch self-reference case.
722 if (L == FnL) {
723 if (R == FnR)
724 return 0;
725 return -1;
726 }
727 if (R == FnR) {
728 if (L == FnL)
729 return 0;
730 return 1;
731 }
732
733 const Constant *ConstL = dyn_cast<Constant>(L);
734 const Constant *ConstR = dyn_cast<Constant>(R);
735 if (ConstL && ConstR) {
736 if (L == R)
737 return 0;
738 return cmpConstants(ConstL, ConstR);
739 }
740
741 if (ConstL)
742 return 1;
743 if (ConstR)
744 return -1;
745
746 const InlineAsm *InlineAsmL = dyn_cast<InlineAsm>(L);
747 const InlineAsm *InlineAsmR = dyn_cast<InlineAsm>(R);
748
749 if (InlineAsmL && InlineAsmR)
750 return cmpInlineAsm(InlineAsmL, InlineAsmR);
751 if (InlineAsmL)
752 return 1;
753 if (InlineAsmR)
754 return -1;
755
756 auto LeftSN = sn_mapL.insert(std::make_pair(L, sn_mapL.size())),
757 RightSN = sn_mapR.insert(std::make_pair(R, sn_mapR.size()));
758
759 return cmpNumbers(LeftSN.first->second, RightSN.first->second);
760 }
761
762 // Test whether two basic blocks have equivalent behaviour.
cmpBasicBlocks(const BasicBlock * BBL,const BasicBlock * BBR) const763 int FunctionComparator::cmpBasicBlocks(const BasicBlock *BBL,
764 const BasicBlock *BBR) const {
765 BasicBlock::const_iterator InstL = BBL->begin(), InstLE = BBL->end();
766 BasicBlock::const_iterator InstR = BBR->begin(), InstRE = BBR->end();
767
768 do {
769 bool needToCmpOperands = true;
770 if (int Res = cmpOperations(&*InstL, &*InstR, needToCmpOperands))
771 return Res;
772 if (needToCmpOperands) {
773 assert(InstL->getNumOperands() == InstR->getNumOperands());
774
775 for (unsigned i = 0, e = InstL->getNumOperands(); i != e; ++i) {
776 Value *OpL = InstL->getOperand(i);
777 Value *OpR = InstR->getOperand(i);
778 if (int Res = cmpValues(OpL, OpR))
779 return Res;
780 // cmpValues should ensure this is true.
781 assert(cmpTypes(OpL->getType(), OpR->getType()) == 0);
782 }
783 }
784
785 ++InstL;
786 ++InstR;
787 } while (InstL != InstLE && InstR != InstRE);
788
789 if (InstL != InstLE && InstR == InstRE)
790 return 1;
791 if (InstL == InstLE && InstR != InstRE)
792 return -1;
793 return 0;
794 }
795
compareSignature() const796 int FunctionComparator::compareSignature() const {
797 if (int Res = cmpAttrs(FnL->getAttributes(), FnR->getAttributes()))
798 return Res;
799
800 if (int Res = cmpNumbers(FnL->hasGC(), FnR->hasGC()))
801 return Res;
802
803 if (FnL->hasGC()) {
804 if (int Res = cmpMem(FnL->getGC(), FnR->getGC()))
805 return Res;
806 }
807
808 if (int Res = cmpNumbers(FnL->hasSection(), FnR->hasSection()))
809 return Res;
810
811 if (FnL->hasSection()) {
812 if (int Res = cmpMem(FnL->getSection(), FnR->getSection()))
813 return Res;
814 }
815
816 if (int Res = cmpNumbers(FnL->isVarArg(), FnR->isVarArg()))
817 return Res;
818
819 // TODO: if it's internal and only used in direct calls, we could handle this
820 // case too.
821 if (int Res = cmpNumbers(FnL->getCallingConv(), FnR->getCallingConv()))
822 return Res;
823
824 if (int Res = cmpTypes(FnL->getFunctionType(), FnR->getFunctionType()))
825 return Res;
826
827 assert(FnL->arg_size() == FnR->arg_size() &&
828 "Identically typed functions have different numbers of args!");
829
830 // Visit the arguments so that they get enumerated in the order they're
831 // passed in.
832 for (Function::const_arg_iterator ArgLI = FnL->arg_begin(),
833 ArgRI = FnR->arg_begin(),
834 ArgLE = FnL->arg_end();
835 ArgLI != ArgLE; ++ArgLI, ++ArgRI) {
836 if (cmpValues(&*ArgLI, &*ArgRI) != 0)
837 llvm_unreachable("Arguments repeat!");
838 }
839 return 0;
840 }
841
842 // Test whether the two functions have equivalent behaviour.
compare()843 int FunctionComparator::compare() {
844 beginCompare();
845
846 if (int Res = compareSignature())
847 return Res;
848
849 // We do a CFG-ordered walk since the actual ordering of the blocks in the
850 // linked list is immaterial. Our walk starts at the entry block for both
851 // functions, then takes each block from each terminator in order. As an
852 // artifact, this also means that unreachable blocks are ignored.
853 SmallVector<const BasicBlock *, 8> FnLBBs, FnRBBs;
854 SmallPtrSet<const BasicBlock *, 32> VisitedBBs; // in terms of F1.
855
856 FnLBBs.push_back(&FnL->getEntryBlock());
857 FnRBBs.push_back(&FnR->getEntryBlock());
858
859 VisitedBBs.insert(FnLBBs[0]);
860 while (!FnLBBs.empty()) {
861 const BasicBlock *BBL = FnLBBs.pop_back_val();
862 const BasicBlock *BBR = FnRBBs.pop_back_val();
863
864 if (int Res = cmpValues(BBL, BBR))
865 return Res;
866
867 if (int Res = cmpBasicBlocks(BBL, BBR))
868 return Res;
869
870 const TerminatorInst *TermL = BBL->getTerminator();
871 const TerminatorInst *TermR = BBR->getTerminator();
872
873 assert(TermL->getNumSuccessors() == TermR->getNumSuccessors());
874 for (unsigned i = 0, e = TermL->getNumSuccessors(); i != e; ++i) {
875 if (!VisitedBBs.insert(TermL->getSuccessor(i)).second)
876 continue;
877
878 FnLBBs.push_back(TermL->getSuccessor(i));
879 FnRBBs.push_back(TermR->getSuccessor(i));
880 }
881 }
882 return 0;
883 }
884
885 namespace {
886
887 // Accumulate the hash of a sequence of 64-bit integers. This is similar to a
888 // hash of a sequence of 64bit ints, but the entire input does not need to be
889 // available at once. This interface is necessary for functionHash because it
890 // needs to accumulate the hash as the structure of the function is traversed
891 // without saving these values to an intermediate buffer. This form of hashing
892 // is not often needed, as usually the object to hash is just read from a
893 // buffer.
894 class HashAccumulator64 {
895 uint64_t Hash;
896
897 public:
898 // Initialize to random constant, so the state isn't zero.
HashAccumulator64()899 HashAccumulator64() { Hash = 0x6acaa36bef8325c5ULL; }
900
add(uint64_t V)901 void add(uint64_t V) {
902 Hash = hashing::detail::hash_16_bytes(Hash, V);
903 }
904
905 // No finishing is required, because the entire hash value is used.
getHash()906 uint64_t getHash() { return Hash; }
907 };
908
909 } // end anonymous namespace
910
911 // A function hash is calculated by considering only the number of arguments and
912 // whether a function is varargs, the order of basic blocks (given by the
913 // successors of each basic block in depth first order), and the order of
914 // opcodes of each instruction within each of these basic blocks. This mirrors
915 // the strategy compare() uses to compare functions by walking the BBs in depth
916 // first order and comparing each instruction in sequence. Because this hash
917 // does not look at the operands, it is insensitive to things such as the
918 // target of calls and the constants used in the function, which makes it useful
919 // when possibly merging functions which are the same modulo constants and call
920 // targets.
functionHash(Function & F)921 FunctionComparator::FunctionHash FunctionComparator::functionHash(Function &F) {
922 HashAccumulator64 H;
923 H.add(F.isVarArg());
924 H.add(F.arg_size());
925
926 SmallVector<const BasicBlock *, 8> BBs;
927 SmallPtrSet<const BasicBlock *, 16> VisitedBBs;
928
929 // Walk the blocks in the same order as FunctionComparator::cmpBasicBlocks(),
930 // accumulating the hash of the function "structure." (BB and opcode sequence)
931 BBs.push_back(&F.getEntryBlock());
932 VisitedBBs.insert(BBs[0]);
933 while (!BBs.empty()) {
934 const BasicBlock *BB = BBs.pop_back_val();
935 // This random value acts as a block header, as otherwise the partition of
936 // opcodes into BBs wouldn't affect the hash, only the order of the opcodes
937 H.add(45798);
938 for (auto &Inst : *BB) {
939 H.add(Inst.getOpcode());
940 }
941 const TerminatorInst *Term = BB->getTerminator();
942 for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) {
943 if (!VisitedBBs.insert(Term->getSuccessor(i)).second)
944 continue;
945 BBs.push_back(Term->getSuccessor(i));
946 }
947 }
948 return H.getHash();
949 }
950