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 case Type::IntegerTyID:
414 return cmpNumbers(cast<IntegerType>(TyL)->getBitWidth(),
415 cast<IntegerType>(TyR)->getBitWidth());
416 // TyL == TyR would have returned true earlier, because types are uniqued.
417 case Type::VoidTyID:
418 case Type::FloatTyID:
419 case Type::DoubleTyID:
420 case Type::X86_FP80TyID:
421 case Type::FP128TyID:
422 case Type::PPC_FP128TyID:
423 case Type::LabelTyID:
424 case Type::MetadataTyID:
425 case Type::TokenTyID:
426 return 0;
427
428 case Type::PointerTyID:
429 assert(PTyL && PTyR && "Both types must be pointers here.");
430 return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
431
432 case Type::StructTyID: {
433 StructType *STyL = cast<StructType>(TyL);
434 StructType *STyR = cast<StructType>(TyR);
435 if (STyL->getNumElements() != STyR->getNumElements())
436 return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
437
438 if (STyL->isPacked() != STyR->isPacked())
439 return cmpNumbers(STyL->isPacked(), STyR->isPacked());
440
441 for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
442 if (int Res = cmpTypes(STyL->getElementType(i), STyR->getElementType(i)))
443 return Res;
444 }
445 return 0;
446 }
447
448 case Type::FunctionTyID: {
449 FunctionType *FTyL = cast<FunctionType>(TyL);
450 FunctionType *FTyR = cast<FunctionType>(TyR);
451 if (FTyL->getNumParams() != FTyR->getNumParams())
452 return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
453
454 if (FTyL->isVarArg() != FTyR->isVarArg())
455 return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
456
457 if (int Res = cmpTypes(FTyL->getReturnType(), FTyR->getReturnType()))
458 return Res;
459
460 for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
461 if (int Res = cmpTypes(FTyL->getParamType(i), FTyR->getParamType(i)))
462 return Res;
463 }
464 return 0;
465 }
466
467 case Type::ArrayTyID:
468 case Type::VectorTyID: {
469 auto *STyL = cast<SequentialType>(TyL);
470 auto *STyR = cast<SequentialType>(TyR);
471 if (STyL->getNumElements() != STyR->getNumElements())
472 return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
473 return cmpTypes(STyL->getElementType(), STyR->getElementType());
474 }
475 }
476 }
477
478 // Determine whether the two operations are the same except that pointer-to-A
479 // and pointer-to-B are equivalent. This should be kept in sync with
480 // Instruction::isSameOperationAs.
481 // Read method declaration comments for more details.
cmpOperations(const Instruction * L,const Instruction * R,bool & needToCmpOperands) const482 int FunctionComparator::cmpOperations(const Instruction *L,
483 const Instruction *R,
484 bool &needToCmpOperands) const {
485 needToCmpOperands = true;
486 if (int Res = cmpValues(L, R))
487 return Res;
488
489 // Differences from Instruction::isSameOperationAs:
490 // * replace type comparison with calls to cmpTypes.
491 // * we test for I->getRawSubclassOptionalData (nuw/nsw/tail) at the top.
492 // * because of the above, we don't test for the tail bit on calls later on.
493 if (int Res = cmpNumbers(L->getOpcode(), R->getOpcode()))
494 return Res;
495
496 if (const GetElementPtrInst *GEPL = dyn_cast<GetElementPtrInst>(L)) {
497 needToCmpOperands = false;
498 const GetElementPtrInst *GEPR = cast<GetElementPtrInst>(R);
499 if (int Res =
500 cmpValues(GEPL->getPointerOperand(), GEPR->getPointerOperand()))
501 return Res;
502 return cmpGEPs(GEPL, GEPR);
503 }
504
505 if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
506 return Res;
507
508 if (int Res = cmpTypes(L->getType(), R->getType()))
509 return Res;
510
511 if (int Res = cmpNumbers(L->getRawSubclassOptionalData(),
512 R->getRawSubclassOptionalData()))
513 return Res;
514
515 // We have two instructions of identical opcode and #operands. Check to see
516 // if all operands are the same type
517 for (unsigned i = 0, e = L->getNumOperands(); i != e; ++i) {
518 if (int Res =
519 cmpTypes(L->getOperand(i)->getType(), R->getOperand(i)->getType()))
520 return Res;
521 }
522
523 // Check special state that is a part of some instructions.
524 if (const AllocaInst *AI = dyn_cast<AllocaInst>(L)) {
525 if (int Res = cmpTypes(AI->getAllocatedType(),
526 cast<AllocaInst>(R)->getAllocatedType()))
527 return Res;
528 return cmpNumbers(AI->getAlignment(), cast<AllocaInst>(R)->getAlignment());
529 }
530 if (const LoadInst *LI = dyn_cast<LoadInst>(L)) {
531 if (int Res = cmpNumbers(LI->isVolatile(), cast<LoadInst>(R)->isVolatile()))
532 return Res;
533 if (int Res =
534 cmpNumbers(LI->getAlignment(), cast<LoadInst>(R)->getAlignment()))
535 return Res;
536 if (int Res =
537 cmpOrderings(LI->getOrdering(), cast<LoadInst>(R)->getOrdering()))
538 return Res;
539 if (int Res = cmpNumbers(LI->getSyncScopeID(),
540 cast<LoadInst>(R)->getSyncScopeID()))
541 return Res;
542 return cmpRangeMetadata(LI->getMetadata(LLVMContext::MD_range),
543 cast<LoadInst>(R)->getMetadata(LLVMContext::MD_range));
544 }
545 if (const StoreInst *SI = dyn_cast<StoreInst>(L)) {
546 if (int Res =
547 cmpNumbers(SI->isVolatile(), cast<StoreInst>(R)->isVolatile()))
548 return Res;
549 if (int Res =
550 cmpNumbers(SI->getAlignment(), cast<StoreInst>(R)->getAlignment()))
551 return Res;
552 if (int Res =
553 cmpOrderings(SI->getOrdering(), cast<StoreInst>(R)->getOrdering()))
554 return Res;
555 return cmpNumbers(SI->getSyncScopeID(),
556 cast<StoreInst>(R)->getSyncScopeID());
557 }
558 if (const CmpInst *CI = dyn_cast<CmpInst>(L))
559 return cmpNumbers(CI->getPredicate(), cast<CmpInst>(R)->getPredicate());
560 if (const CallInst *CI = dyn_cast<CallInst>(L)) {
561 if (int Res = cmpNumbers(CI->getCallingConv(),
562 cast<CallInst>(R)->getCallingConv()))
563 return Res;
564 if (int Res =
565 cmpAttrs(CI->getAttributes(), cast<CallInst>(R)->getAttributes()))
566 return Res;
567 if (int Res = cmpOperandBundlesSchema(CI, R))
568 return Res;
569 return cmpRangeMetadata(
570 CI->getMetadata(LLVMContext::MD_range),
571 cast<CallInst>(R)->getMetadata(LLVMContext::MD_range));
572 }
573 if (const InvokeInst *II = dyn_cast<InvokeInst>(L)) {
574 if (int Res = cmpNumbers(II->getCallingConv(),
575 cast<InvokeInst>(R)->getCallingConv()))
576 return Res;
577 if (int Res =
578 cmpAttrs(II->getAttributes(), cast<InvokeInst>(R)->getAttributes()))
579 return Res;
580 if (int Res = cmpOperandBundlesSchema(II, R))
581 return Res;
582 return cmpRangeMetadata(
583 II->getMetadata(LLVMContext::MD_range),
584 cast<InvokeInst>(R)->getMetadata(LLVMContext::MD_range));
585 }
586 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(L)) {
587 ArrayRef<unsigned> LIndices = IVI->getIndices();
588 ArrayRef<unsigned> RIndices = cast<InsertValueInst>(R)->getIndices();
589 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
590 return Res;
591 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
592 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
593 return Res;
594 }
595 return 0;
596 }
597 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(L)) {
598 ArrayRef<unsigned> LIndices = EVI->getIndices();
599 ArrayRef<unsigned> RIndices = cast<ExtractValueInst>(R)->getIndices();
600 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
601 return Res;
602 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
603 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
604 return Res;
605 }
606 }
607 if (const FenceInst *FI = dyn_cast<FenceInst>(L)) {
608 if (int Res =
609 cmpOrderings(FI->getOrdering(), cast<FenceInst>(R)->getOrdering()))
610 return Res;
611 return cmpNumbers(FI->getSyncScopeID(),
612 cast<FenceInst>(R)->getSyncScopeID());
613 }
614 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(L)) {
615 if (int Res = cmpNumbers(CXI->isVolatile(),
616 cast<AtomicCmpXchgInst>(R)->isVolatile()))
617 return Res;
618 if (int Res = cmpNumbers(CXI->isWeak(),
619 cast<AtomicCmpXchgInst>(R)->isWeak()))
620 return Res;
621 if (int Res =
622 cmpOrderings(CXI->getSuccessOrdering(),
623 cast<AtomicCmpXchgInst>(R)->getSuccessOrdering()))
624 return Res;
625 if (int Res =
626 cmpOrderings(CXI->getFailureOrdering(),
627 cast<AtomicCmpXchgInst>(R)->getFailureOrdering()))
628 return Res;
629 return cmpNumbers(CXI->getSyncScopeID(),
630 cast<AtomicCmpXchgInst>(R)->getSyncScopeID());
631 }
632 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(L)) {
633 if (int Res = cmpNumbers(RMWI->getOperation(),
634 cast<AtomicRMWInst>(R)->getOperation()))
635 return Res;
636 if (int Res = cmpNumbers(RMWI->isVolatile(),
637 cast<AtomicRMWInst>(R)->isVolatile()))
638 return Res;
639 if (int Res = cmpOrderings(RMWI->getOrdering(),
640 cast<AtomicRMWInst>(R)->getOrdering()))
641 return Res;
642 return cmpNumbers(RMWI->getSyncScopeID(),
643 cast<AtomicRMWInst>(R)->getSyncScopeID());
644 }
645 if (const PHINode *PNL = dyn_cast<PHINode>(L)) {
646 const PHINode *PNR = cast<PHINode>(R);
647 // Ensure that in addition to the incoming values being identical
648 // (checked by the caller of this function), the incoming blocks
649 // are also identical.
650 for (unsigned i = 0, e = PNL->getNumIncomingValues(); i != e; ++i) {
651 if (int Res =
652 cmpValues(PNL->getIncomingBlock(i), PNR->getIncomingBlock(i)))
653 return Res;
654 }
655 }
656 return 0;
657 }
658
659 // Determine whether two GEP operations perform the same underlying arithmetic.
660 // Read method declaration comments for more details.
cmpGEPs(const GEPOperator * GEPL,const GEPOperator * GEPR) const661 int FunctionComparator::cmpGEPs(const GEPOperator *GEPL,
662 const GEPOperator *GEPR) const {
663 unsigned int ASL = GEPL->getPointerAddressSpace();
664 unsigned int ASR = GEPR->getPointerAddressSpace();
665
666 if (int Res = cmpNumbers(ASL, ASR))
667 return Res;
668
669 // When we have target data, we can reduce the GEP down to the value in bytes
670 // added to the address.
671 const DataLayout &DL = FnL->getParent()->getDataLayout();
672 unsigned BitWidth = DL.getPointerSizeInBits(ASL);
673 APInt OffsetL(BitWidth, 0), OffsetR(BitWidth, 0);
674 if (GEPL->accumulateConstantOffset(DL, OffsetL) &&
675 GEPR->accumulateConstantOffset(DL, OffsetR))
676 return cmpAPInts(OffsetL, OffsetR);
677 if (int Res = cmpTypes(GEPL->getSourceElementType(),
678 GEPR->getSourceElementType()))
679 return Res;
680
681 if (int Res = cmpNumbers(GEPL->getNumOperands(), GEPR->getNumOperands()))
682 return Res;
683
684 for (unsigned i = 0, e = GEPL->getNumOperands(); i != e; ++i) {
685 if (int Res = cmpValues(GEPL->getOperand(i), GEPR->getOperand(i)))
686 return Res;
687 }
688
689 return 0;
690 }
691
cmpInlineAsm(const InlineAsm * L,const InlineAsm * R) const692 int FunctionComparator::cmpInlineAsm(const InlineAsm *L,
693 const InlineAsm *R) const {
694 // InlineAsm's are uniqued. If they are the same pointer, obviously they are
695 // the same, otherwise compare the fields.
696 if (L == R)
697 return 0;
698 if (int Res = cmpTypes(L->getFunctionType(), R->getFunctionType()))
699 return Res;
700 if (int Res = cmpMem(L->getAsmString(), R->getAsmString()))
701 return Res;
702 if (int Res = cmpMem(L->getConstraintString(), R->getConstraintString()))
703 return Res;
704 if (int Res = cmpNumbers(L->hasSideEffects(), R->hasSideEffects()))
705 return Res;
706 if (int Res = cmpNumbers(L->isAlignStack(), R->isAlignStack()))
707 return Res;
708 if (int Res = cmpNumbers(L->getDialect(), R->getDialect()))
709 return Res;
710 assert(L->getFunctionType() != R->getFunctionType());
711 return 0;
712 }
713
714 /// Compare two values used by the two functions under pair-wise comparison. If
715 /// this is the first time the values are seen, they're added to the mapping so
716 /// that we will detect mismatches on next use.
717 /// See comments in declaration for more details.
cmpValues(const Value * L,const Value * R) const718 int FunctionComparator::cmpValues(const Value *L, const Value *R) const {
719 // Catch self-reference case.
720 if (L == FnL) {
721 if (R == FnR)
722 return 0;
723 return -1;
724 }
725 if (R == FnR) {
726 if (L == FnL)
727 return 0;
728 return 1;
729 }
730
731 const Constant *ConstL = dyn_cast<Constant>(L);
732 const Constant *ConstR = dyn_cast<Constant>(R);
733 if (ConstL && ConstR) {
734 if (L == R)
735 return 0;
736 return cmpConstants(ConstL, ConstR);
737 }
738
739 if (ConstL)
740 return 1;
741 if (ConstR)
742 return -1;
743
744 const InlineAsm *InlineAsmL = dyn_cast<InlineAsm>(L);
745 const InlineAsm *InlineAsmR = dyn_cast<InlineAsm>(R);
746
747 if (InlineAsmL && InlineAsmR)
748 return cmpInlineAsm(InlineAsmL, InlineAsmR);
749 if (InlineAsmL)
750 return 1;
751 if (InlineAsmR)
752 return -1;
753
754 auto LeftSN = sn_mapL.insert(std::make_pair(L, sn_mapL.size())),
755 RightSN = sn_mapR.insert(std::make_pair(R, sn_mapR.size()));
756
757 return cmpNumbers(LeftSN.first->second, RightSN.first->second);
758 }
759
760 // Test whether two basic blocks have equivalent behaviour.
cmpBasicBlocks(const BasicBlock * BBL,const BasicBlock * BBR) const761 int FunctionComparator::cmpBasicBlocks(const BasicBlock *BBL,
762 const BasicBlock *BBR) const {
763 BasicBlock::const_iterator InstL = BBL->begin(), InstLE = BBL->end();
764 BasicBlock::const_iterator InstR = BBR->begin(), InstRE = BBR->end();
765
766 do {
767 bool needToCmpOperands = true;
768 if (int Res = cmpOperations(&*InstL, &*InstR, needToCmpOperands))
769 return Res;
770 if (needToCmpOperands) {
771 assert(InstL->getNumOperands() == InstR->getNumOperands());
772
773 for (unsigned i = 0, e = InstL->getNumOperands(); i != e; ++i) {
774 Value *OpL = InstL->getOperand(i);
775 Value *OpR = InstR->getOperand(i);
776 if (int Res = cmpValues(OpL, OpR))
777 return Res;
778 // cmpValues should ensure this is true.
779 assert(cmpTypes(OpL->getType(), OpR->getType()) == 0);
780 }
781 }
782
783 ++InstL;
784 ++InstR;
785 } while (InstL != InstLE && InstR != InstRE);
786
787 if (InstL != InstLE && InstR == InstRE)
788 return 1;
789 if (InstL == InstLE && InstR != InstRE)
790 return -1;
791 return 0;
792 }
793
compareSignature() const794 int FunctionComparator::compareSignature() const {
795 if (int Res = cmpAttrs(FnL->getAttributes(), FnR->getAttributes()))
796 return Res;
797
798 if (int Res = cmpNumbers(FnL->hasGC(), FnR->hasGC()))
799 return Res;
800
801 if (FnL->hasGC()) {
802 if (int Res = cmpMem(FnL->getGC(), FnR->getGC()))
803 return Res;
804 }
805
806 if (int Res = cmpNumbers(FnL->hasSection(), FnR->hasSection()))
807 return Res;
808
809 if (FnL->hasSection()) {
810 if (int Res = cmpMem(FnL->getSection(), FnR->getSection()))
811 return Res;
812 }
813
814 if (int Res = cmpNumbers(FnL->isVarArg(), FnR->isVarArg()))
815 return Res;
816
817 // TODO: if it's internal and only used in direct calls, we could handle this
818 // case too.
819 if (int Res = cmpNumbers(FnL->getCallingConv(), FnR->getCallingConv()))
820 return Res;
821
822 if (int Res = cmpTypes(FnL->getFunctionType(), FnR->getFunctionType()))
823 return Res;
824
825 assert(FnL->arg_size() == FnR->arg_size() &&
826 "Identically typed functions have different numbers of args!");
827
828 // Visit the arguments so that they get enumerated in the order they're
829 // passed in.
830 for (Function::const_arg_iterator ArgLI = FnL->arg_begin(),
831 ArgRI = FnR->arg_begin(),
832 ArgLE = FnL->arg_end();
833 ArgLI != ArgLE; ++ArgLI, ++ArgRI) {
834 if (cmpValues(&*ArgLI, &*ArgRI) != 0)
835 llvm_unreachable("Arguments repeat!");
836 }
837 return 0;
838 }
839
840 // Test whether the two functions have equivalent behaviour.
compare()841 int FunctionComparator::compare() {
842 beginCompare();
843
844 if (int Res = compareSignature())
845 return Res;
846
847 // We do a CFG-ordered walk since the actual ordering of the blocks in the
848 // linked list is immaterial. Our walk starts at the entry block for both
849 // functions, then takes each block from each terminator in order. As an
850 // artifact, this also means that unreachable blocks are ignored.
851 SmallVector<const BasicBlock *, 8> FnLBBs, FnRBBs;
852 SmallPtrSet<const BasicBlock *, 32> VisitedBBs; // in terms of F1.
853
854 FnLBBs.push_back(&FnL->getEntryBlock());
855 FnRBBs.push_back(&FnR->getEntryBlock());
856
857 VisitedBBs.insert(FnLBBs[0]);
858 while (!FnLBBs.empty()) {
859 const BasicBlock *BBL = FnLBBs.pop_back_val();
860 const BasicBlock *BBR = FnRBBs.pop_back_val();
861
862 if (int Res = cmpValues(BBL, BBR))
863 return Res;
864
865 if (int Res = cmpBasicBlocks(BBL, BBR))
866 return Res;
867
868 const Instruction *TermL = BBL->getTerminator();
869 const Instruction *TermR = BBR->getTerminator();
870
871 assert(TermL->getNumSuccessors() == TermR->getNumSuccessors());
872 for (unsigned i = 0, e = TermL->getNumSuccessors(); i != e; ++i) {
873 if (!VisitedBBs.insert(TermL->getSuccessor(i)).second)
874 continue;
875
876 FnLBBs.push_back(TermL->getSuccessor(i));
877 FnRBBs.push_back(TermR->getSuccessor(i));
878 }
879 }
880 return 0;
881 }
882
883 namespace {
884
885 // Accumulate the hash of a sequence of 64-bit integers. This is similar to a
886 // hash of a sequence of 64bit ints, but the entire input does not need to be
887 // available at once. This interface is necessary for functionHash because it
888 // needs to accumulate the hash as the structure of the function is traversed
889 // without saving these values to an intermediate buffer. This form of hashing
890 // is not often needed, as usually the object to hash is just read from a
891 // buffer.
892 class HashAccumulator64 {
893 uint64_t Hash;
894
895 public:
896 // Initialize to random constant, so the state isn't zero.
HashAccumulator64()897 HashAccumulator64() { Hash = 0x6acaa36bef8325c5ULL; }
898
add(uint64_t V)899 void add(uint64_t V) {
900 Hash = hashing::detail::hash_16_bytes(Hash, V);
901 }
902
903 // No finishing is required, because the entire hash value is used.
getHash()904 uint64_t getHash() { return Hash; }
905 };
906
907 } // end anonymous namespace
908
909 // A function hash is calculated by considering only the number of arguments and
910 // whether a function is varargs, the order of basic blocks (given by the
911 // successors of each basic block in depth first order), and the order of
912 // opcodes of each instruction within each of these basic blocks. This mirrors
913 // the strategy compare() uses to compare functions by walking the BBs in depth
914 // first order and comparing each instruction in sequence. Because this hash
915 // does not look at the operands, it is insensitive to things such as the
916 // target of calls and the constants used in the function, which makes it useful
917 // when possibly merging functions which are the same modulo constants and call
918 // targets.
functionHash(Function & F)919 FunctionComparator::FunctionHash FunctionComparator::functionHash(Function &F) {
920 HashAccumulator64 H;
921 H.add(F.isVarArg());
922 H.add(F.arg_size());
923
924 SmallVector<const BasicBlock *, 8> BBs;
925 SmallPtrSet<const BasicBlock *, 16> VisitedBBs;
926
927 // Walk the blocks in the same order as FunctionComparator::cmpBasicBlocks(),
928 // accumulating the hash of the function "structure." (BB and opcode sequence)
929 BBs.push_back(&F.getEntryBlock());
930 VisitedBBs.insert(BBs[0]);
931 while (!BBs.empty()) {
932 const BasicBlock *BB = BBs.pop_back_val();
933 // This random value acts as a block header, as otherwise the partition of
934 // opcodes into BBs wouldn't affect the hash, only the order of the opcodes
935 H.add(45798);
936 for (auto &Inst : *BB) {
937 H.add(Inst.getOpcode());
938 }
939 const Instruction *Term = BB->getTerminator();
940 for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) {
941 if (!VisitedBBs.insert(Term->getSuccessor(i)).second)
942 continue;
943 BBs.push_back(Term->getSuccessor(i));
944 }
945 }
946 return H.getHash();
947 }
948