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/Constant.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GlobalValue.h"
29 #include "llvm/IR/InlineAsm.h"
30 #include "llvm/IR/InstrTypes.h"
31 #include "llvm/IR/Instruction.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/LLVMContext.h"
34 #include "llvm/IR/Metadata.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/IR/Type.h"
38 #include "llvm/IR/Value.h"
39 #include "llvm/Support/Casting.h"
40 #include "llvm/Support/Compiler.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include <cassert>
45 #include <cstddef>
46 #include <cstdint>
47 #include <utility>
48
49 using namespace llvm;
50
51 #define DEBUG_TYPE "functioncomparator"
52
cmpNumbers(uint64_t L,uint64_t R) const53 int FunctionComparator::cmpNumbers(uint64_t L, uint64_t R) const {
54 if (L < R)
55 return -1;
56 if (L > R)
57 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)
63 return -1;
64 if ((int)L > (int)R)
65 return 1;
66 return 0;
67 }
68
cmpAPInts(const APInt & L,const APInt & R) const69 int FunctionComparator::cmpAPInts(const APInt &L, const APInt &R) const {
70 if (int Res = cmpNumbers(L.getBitWidth(), R.getBitWidth()))
71 return Res;
72 if (L.ugt(R))
73 return 1;
74 if (R.ugt(L))
75 return -1;
76 return 0;
77 }
78
cmpAPFloats(const APFloat & L,const APFloat & R) const79 int FunctionComparator::cmpAPFloats(const APFloat &L, const APFloat &R) const {
80 // Floats are ordered first by semantics (i.e. float, double, half, etc.),
81 // then by value interpreted as a bitstring (aka APInt).
82 const fltSemantics &SL = L.getSemantics(), &SR = R.getSemantics();
83 if (int Res = cmpNumbers(APFloat::semanticsPrecision(SL),
84 APFloat::semanticsPrecision(SR)))
85 return Res;
86 if (int Res = cmpNumbers(APFloat::semanticsMaxExponent(SL),
87 APFloat::semanticsMaxExponent(SR)))
88 return Res;
89 if (int Res = cmpNumbers(APFloat::semanticsMinExponent(SL),
90 APFloat::semanticsMinExponent(SR)))
91 return Res;
92 if (int Res = cmpNumbers(APFloat::semanticsSizeInBits(SL),
93 APFloat::semanticsSizeInBits(SR)))
94 return Res;
95 return cmpAPInts(L.bitcastToAPInt(), R.bitcastToAPInt());
96 }
97
cmpMem(StringRef L,StringRef R) const98 int FunctionComparator::cmpMem(StringRef L, StringRef R) const {
99 // Prevent heavy comparison, compare sizes first.
100 if (int Res = cmpNumbers(L.size(), R.size()))
101 return Res;
102
103 // Compare strings lexicographically only when it is necessary: only when
104 // strings are equal in size.
105 return L.compare(R);
106 }
107
cmpAttrs(const AttributeList L,const AttributeList R) const108 int FunctionComparator::cmpAttrs(const AttributeList L,
109 const AttributeList R) const {
110 if (int Res = cmpNumbers(L.getNumAttrSets(), R.getNumAttrSets()))
111 return Res;
112
113 for (unsigned i = L.index_begin(), e = L.index_end(); i != e; ++i) {
114 AttributeSet LAS = L.getAttributes(i);
115 AttributeSet RAS = R.getAttributes(i);
116 AttributeSet::iterator LI = LAS.begin(), LE = LAS.end();
117 AttributeSet::iterator RI = RAS.begin(), RE = RAS.end();
118 for (; LI != LE && RI != RE; ++LI, ++RI) {
119 Attribute LA = *LI;
120 Attribute RA = *RI;
121 if (LA.isTypeAttribute() && RA.isTypeAttribute()) {
122 if (LA.getKindAsEnum() != RA.getKindAsEnum())
123 return cmpNumbers(LA.getKindAsEnum(), RA.getKindAsEnum());
124
125 Type *TyL = LA.getValueAsType();
126 Type *TyR = RA.getValueAsType();
127 if (TyL && TyR) {
128 if (int Res = cmpTypes(TyL, TyR))
129 return Res;
130 continue;
131 }
132
133 // Two pointers, at least one null, so the comparison result is
134 // independent of the value of a real pointer.
135 if (int Res = cmpNumbers((uint64_t)TyL, (uint64_t)TyR))
136 return Res;
137 continue;
138 }
139 if (LA < RA)
140 return -1;
141 if (RA < LA)
142 return 1;
143 }
144 if (LI != LE)
145 return 1;
146 if (RI != RE)
147 return -1;
148 }
149 return 0;
150 }
151
cmpRangeMetadata(const MDNode * L,const MDNode * R) const152 int FunctionComparator::cmpRangeMetadata(const MDNode *L,
153 const MDNode *R) const {
154 if (L == R)
155 return 0;
156 if (!L)
157 return -1;
158 if (!R)
159 return 1;
160 // Range metadata is a sequence of numbers. Make sure they are the same
161 // sequence.
162 // TODO: Note that as this is metadata, it is possible to drop and/or merge
163 // this data when considering functions to merge. Thus this comparison would
164 // return 0 (i.e. equivalent), but merging would become more complicated
165 // because the ranges would need to be unioned. It is not likely that
166 // functions differ ONLY in this metadata if they are actually the same
167 // function semantically.
168 if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
169 return Res;
170 for (size_t I = 0; I < L->getNumOperands(); ++I) {
171 ConstantInt *LLow = mdconst::extract<ConstantInt>(L->getOperand(I));
172 ConstantInt *RLow = mdconst::extract<ConstantInt>(R->getOperand(I));
173 if (int Res = cmpAPInts(LLow->getValue(), RLow->getValue()))
174 return Res;
175 }
176 return 0;
177 }
178
cmpOperandBundlesSchema(const CallBase & LCS,const CallBase & RCS) const179 int FunctionComparator::cmpOperandBundlesSchema(const CallBase &LCS,
180 const CallBase &RCS) const {
181 assert(LCS.getOpcode() == RCS.getOpcode() && "Can't compare otherwise!");
182
183 if (int Res =
184 cmpNumbers(LCS.getNumOperandBundles(), RCS.getNumOperandBundles()))
185 return Res;
186
187 for (unsigned I = 0, E = LCS.getNumOperandBundles(); I != E; ++I) {
188 auto OBL = LCS.getOperandBundleAt(I);
189 auto OBR = RCS.getOperandBundleAt(I);
190
191 if (int Res = OBL.getTagName().compare(OBR.getTagName()))
192 return Res;
193
194 if (int Res = cmpNumbers(OBL.Inputs.size(), OBR.Inputs.size()))
195 return Res;
196 }
197
198 return 0;
199 }
200
201 /// Constants comparison:
202 /// 1. Check whether type of L constant could be losslessly bitcasted to R
203 /// type.
204 /// 2. Compare constant contents.
205 /// For more details see declaration comments.
cmpConstants(const Constant * L,const Constant * R) const206 int FunctionComparator::cmpConstants(const Constant *L,
207 const Constant *R) const {
208 Type *TyL = L->getType();
209 Type *TyR = R->getType();
210
211 // Check whether types are bitcastable. This part is just re-factored
212 // Type::canLosslesslyBitCastTo method, but instead of returning true/false,
213 // we also pack into result which type is "less" for us.
214 int TypesRes = cmpTypes(TyL, TyR);
215 if (TypesRes != 0) {
216 // Types are different, but check whether we can bitcast them.
217 if (!TyL->isFirstClassType()) {
218 if (TyR->isFirstClassType())
219 return -1;
220 // Neither TyL nor TyR are values of first class type. Return the result
221 // of comparing the types
222 return TypesRes;
223 }
224 if (!TyR->isFirstClassType()) {
225 if (TyL->isFirstClassType())
226 return 1;
227 return TypesRes;
228 }
229
230 // Vector -> Vector conversions are always lossless if the two vector types
231 // have the same size, otherwise not.
232 unsigned TyLWidth = 0;
233 unsigned TyRWidth = 0;
234
235 if (auto *VecTyL = dyn_cast<VectorType>(TyL))
236 TyLWidth = VecTyL->getPrimitiveSizeInBits().getFixedSize();
237 if (auto *VecTyR = dyn_cast<VectorType>(TyR))
238 TyRWidth = VecTyR->getPrimitiveSizeInBits().getFixedSize();
239
240 if (TyLWidth != TyRWidth)
241 return cmpNumbers(TyLWidth, TyRWidth);
242
243 // Zero bit-width means neither TyL nor TyR are vectors.
244 if (!TyLWidth) {
245 PointerType *PTyL = dyn_cast<PointerType>(TyL);
246 PointerType *PTyR = dyn_cast<PointerType>(TyR);
247 if (PTyL && PTyR) {
248 unsigned AddrSpaceL = PTyL->getAddressSpace();
249 unsigned AddrSpaceR = PTyR->getAddressSpace();
250 if (int Res = cmpNumbers(AddrSpaceL, AddrSpaceR))
251 return Res;
252 }
253 if (PTyL)
254 return 1;
255 if (PTyR)
256 return -1;
257
258 // TyL and TyR aren't vectors, nor pointers. We don't know how to
259 // bitcast them.
260 return TypesRes;
261 }
262 }
263
264 // OK, types are bitcastable, now check constant contents.
265
266 if (L->isNullValue() && R->isNullValue())
267 return TypesRes;
268 if (L->isNullValue() && !R->isNullValue())
269 return 1;
270 if (!L->isNullValue() && R->isNullValue())
271 return -1;
272
273 auto GlobalValueL = const_cast<GlobalValue *>(dyn_cast<GlobalValue>(L));
274 auto GlobalValueR = const_cast<GlobalValue *>(dyn_cast<GlobalValue>(R));
275 if (GlobalValueL && GlobalValueR) {
276 return cmpGlobalValues(GlobalValueL, GlobalValueR);
277 }
278
279 if (int Res = cmpNumbers(L->getValueID(), R->getValueID()))
280 return Res;
281
282 if (const auto *SeqL = dyn_cast<ConstantDataSequential>(L)) {
283 const auto *SeqR = cast<ConstantDataSequential>(R);
284 // This handles ConstantDataArray and ConstantDataVector. Note that we
285 // compare the two raw data arrays, which might differ depending on the host
286 // endianness. This isn't a problem though, because the endiness of a module
287 // will affect the order of the constants, but this order is the same
288 // for a given input module and host platform.
289 return cmpMem(SeqL->getRawDataValues(), SeqR->getRawDataValues());
290 }
291
292 switch (L->getValueID()) {
293 case Value::UndefValueVal:
294 case Value::PoisonValueVal:
295 case Value::ConstantTokenNoneVal:
296 return TypesRes;
297 case Value::ConstantIntVal: {
298 const APInt &LInt = cast<ConstantInt>(L)->getValue();
299 const APInt &RInt = cast<ConstantInt>(R)->getValue();
300 return cmpAPInts(LInt, RInt);
301 }
302 case Value::ConstantFPVal: {
303 const APFloat &LAPF = cast<ConstantFP>(L)->getValueAPF();
304 const APFloat &RAPF = cast<ConstantFP>(R)->getValueAPF();
305 return cmpAPFloats(LAPF, RAPF);
306 }
307 case Value::ConstantArrayVal: {
308 const ConstantArray *LA = cast<ConstantArray>(L);
309 const ConstantArray *RA = cast<ConstantArray>(R);
310 uint64_t NumElementsL = cast<ArrayType>(TyL)->getNumElements();
311 uint64_t NumElementsR = cast<ArrayType>(TyR)->getNumElements();
312 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
313 return Res;
314 for (uint64_t i = 0; i < NumElementsL; ++i) {
315 if (int Res = cmpConstants(cast<Constant>(LA->getOperand(i)),
316 cast<Constant>(RA->getOperand(i))))
317 return Res;
318 }
319 return 0;
320 }
321 case Value::ConstantStructVal: {
322 const ConstantStruct *LS = cast<ConstantStruct>(L);
323 const ConstantStruct *RS = cast<ConstantStruct>(R);
324 unsigned NumElementsL = cast<StructType>(TyL)->getNumElements();
325 unsigned NumElementsR = cast<StructType>(TyR)->getNumElements();
326 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
327 return Res;
328 for (unsigned i = 0; i != NumElementsL; ++i) {
329 if (int Res = cmpConstants(cast<Constant>(LS->getOperand(i)),
330 cast<Constant>(RS->getOperand(i))))
331 return Res;
332 }
333 return 0;
334 }
335 case Value::ConstantVectorVal: {
336 const ConstantVector *LV = cast<ConstantVector>(L);
337 const ConstantVector *RV = cast<ConstantVector>(R);
338 unsigned NumElementsL = cast<FixedVectorType>(TyL)->getNumElements();
339 unsigned NumElementsR = cast<FixedVectorType>(TyR)->getNumElements();
340 if (int Res = cmpNumbers(NumElementsL, NumElementsR))
341 return Res;
342 for (uint64_t i = 0; i < NumElementsL; ++i) {
343 if (int Res = cmpConstants(cast<Constant>(LV->getOperand(i)),
344 cast<Constant>(RV->getOperand(i))))
345 return Res;
346 }
347 return 0;
348 }
349 case Value::ConstantExprVal: {
350 const ConstantExpr *LE = cast<ConstantExpr>(L);
351 const ConstantExpr *RE = cast<ConstantExpr>(R);
352 unsigned NumOperandsL = LE->getNumOperands();
353 unsigned NumOperandsR = RE->getNumOperands();
354 if (int Res = cmpNumbers(NumOperandsL, NumOperandsR))
355 return Res;
356 for (unsigned i = 0; i < NumOperandsL; ++i) {
357 if (int Res = cmpConstants(cast<Constant>(LE->getOperand(i)),
358 cast<Constant>(RE->getOperand(i))))
359 return Res;
360 }
361 return 0;
362 }
363 case Value::BlockAddressVal: {
364 const BlockAddress *LBA = cast<BlockAddress>(L);
365 const BlockAddress *RBA = cast<BlockAddress>(R);
366 if (int Res = cmpValues(LBA->getFunction(), RBA->getFunction()))
367 return Res;
368 if (LBA->getFunction() == RBA->getFunction()) {
369 // They are BBs in the same function. Order by which comes first in the
370 // BB order of the function. This order is deterministic.
371 Function *F = LBA->getFunction();
372 BasicBlock *LBB = LBA->getBasicBlock();
373 BasicBlock *RBB = RBA->getBasicBlock();
374 if (LBB == RBB)
375 return 0;
376 for (BasicBlock &BB : F->getBasicBlockList()) {
377 if (&BB == LBB) {
378 assert(&BB != RBB);
379 return -1;
380 }
381 if (&BB == RBB)
382 return 1;
383 }
384 llvm_unreachable("Basic Block Address does not point to a basic block in "
385 "its function.");
386 return -1;
387 } else {
388 // cmpValues said the functions are the same. So because they aren't
389 // literally the same pointer, they must respectively be the left and
390 // right functions.
391 assert(LBA->getFunction() == FnL && RBA->getFunction() == FnR);
392 // cmpValues will tell us if these are equivalent BasicBlocks, in the
393 // context of their respective functions.
394 return cmpValues(LBA->getBasicBlock(), RBA->getBasicBlock());
395 }
396 }
397 default: // Unknown constant, abort.
398 LLVM_DEBUG(dbgs() << "Looking at valueID " << L->getValueID() << "\n");
399 llvm_unreachable("Constant ValueID not recognized.");
400 return -1;
401 }
402 }
403
cmpGlobalValues(GlobalValue * L,GlobalValue * R) const404 int FunctionComparator::cmpGlobalValues(GlobalValue *L, GlobalValue *R) const {
405 uint64_t LNumber = GlobalNumbers->getNumber(L);
406 uint64_t RNumber = GlobalNumbers->getNumber(R);
407 return cmpNumbers(LNumber, RNumber);
408 }
409
410 /// cmpType - compares two types,
411 /// defines total ordering among the types set.
412 /// See method declaration comments for more details.
cmpTypes(Type * TyL,Type * TyR) const413 int FunctionComparator::cmpTypes(Type *TyL, Type *TyR) const {
414 PointerType *PTyL = dyn_cast<PointerType>(TyL);
415 PointerType *PTyR = dyn_cast<PointerType>(TyR);
416
417 const DataLayout &DL = FnL->getParent()->getDataLayout();
418 if (PTyL && PTyL->getAddressSpace() == 0)
419 TyL = DL.getIntPtrType(TyL);
420 if (PTyR && PTyR->getAddressSpace() == 0)
421 TyR = DL.getIntPtrType(TyR);
422
423 if (TyL == TyR)
424 return 0;
425
426 if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
427 return Res;
428
429 switch (TyL->getTypeID()) {
430 default:
431 llvm_unreachable("Unknown type!");
432 case Type::IntegerTyID:
433 return cmpNumbers(cast<IntegerType>(TyL)->getBitWidth(),
434 cast<IntegerType>(TyR)->getBitWidth());
435 // TyL == TyR would have returned true earlier, because types are uniqued.
436 case Type::VoidTyID:
437 case Type::FloatTyID:
438 case Type::DoubleTyID:
439 case Type::X86_FP80TyID:
440 case Type::FP128TyID:
441 case Type::PPC_FP128TyID:
442 case Type::LabelTyID:
443 case Type::MetadataTyID:
444 case Type::TokenTyID:
445 return 0;
446
447 case Type::PointerTyID:
448 assert(PTyL && PTyR && "Both types must be pointers here.");
449 return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
450
451 case Type::StructTyID: {
452 StructType *STyL = cast<StructType>(TyL);
453 StructType *STyR = cast<StructType>(TyR);
454 if (STyL->getNumElements() != STyR->getNumElements())
455 return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
456
457 if (STyL->isPacked() != STyR->isPacked())
458 return cmpNumbers(STyL->isPacked(), STyR->isPacked());
459
460 for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
461 if (int Res = cmpTypes(STyL->getElementType(i), STyR->getElementType(i)))
462 return Res;
463 }
464 return 0;
465 }
466
467 case Type::FunctionTyID: {
468 FunctionType *FTyL = cast<FunctionType>(TyL);
469 FunctionType *FTyR = cast<FunctionType>(TyR);
470 if (FTyL->getNumParams() != FTyR->getNumParams())
471 return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
472
473 if (FTyL->isVarArg() != FTyR->isVarArg())
474 return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
475
476 if (int Res = cmpTypes(FTyL->getReturnType(), FTyR->getReturnType()))
477 return Res;
478
479 for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
480 if (int Res = cmpTypes(FTyL->getParamType(i), FTyR->getParamType(i)))
481 return Res;
482 }
483 return 0;
484 }
485
486 case Type::ArrayTyID: {
487 auto *STyL = cast<ArrayType>(TyL);
488 auto *STyR = cast<ArrayType>(TyR);
489 if (STyL->getNumElements() != STyR->getNumElements())
490 return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
491 return cmpTypes(STyL->getElementType(), STyR->getElementType());
492 }
493 case Type::FixedVectorTyID:
494 case Type::ScalableVectorTyID: {
495 auto *STyL = cast<VectorType>(TyL);
496 auto *STyR = cast<VectorType>(TyR);
497 if (STyL->getElementCount().isScalable() !=
498 STyR->getElementCount().isScalable())
499 return cmpNumbers(STyL->getElementCount().isScalable(),
500 STyR->getElementCount().isScalable());
501 if (STyL->getElementCount() != STyR->getElementCount())
502 return cmpNumbers(STyL->getElementCount().getKnownMinValue(),
503 STyR->getElementCount().getKnownMinValue());
504 return cmpTypes(STyL->getElementType(), STyR->getElementType());
505 }
506 }
507 }
508
509 // Determine whether the two operations are the same except that pointer-to-A
510 // and pointer-to-B are equivalent. This should be kept in sync with
511 // Instruction::isSameOperationAs.
512 // Read method declaration comments for more details.
cmpOperations(const Instruction * L,const Instruction * R,bool & needToCmpOperands) const513 int FunctionComparator::cmpOperations(const Instruction *L,
514 const Instruction *R,
515 bool &needToCmpOperands) const {
516 needToCmpOperands = true;
517 if (int Res = cmpValues(L, R))
518 return Res;
519
520 // Differences from Instruction::isSameOperationAs:
521 // * replace type comparison with calls to cmpTypes.
522 // * we test for I->getRawSubclassOptionalData (nuw/nsw/tail) at the top.
523 // * because of the above, we don't test for the tail bit on calls later on.
524 if (int Res = cmpNumbers(L->getOpcode(), R->getOpcode()))
525 return Res;
526
527 if (const GetElementPtrInst *GEPL = dyn_cast<GetElementPtrInst>(L)) {
528 needToCmpOperands = false;
529 const GetElementPtrInst *GEPR = cast<GetElementPtrInst>(R);
530 if (int Res =
531 cmpValues(GEPL->getPointerOperand(), GEPR->getPointerOperand()))
532 return Res;
533 return cmpGEPs(GEPL, GEPR);
534 }
535
536 if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
537 return Res;
538
539 if (int Res = cmpTypes(L->getType(), R->getType()))
540 return Res;
541
542 if (int Res = cmpNumbers(L->getRawSubclassOptionalData(),
543 R->getRawSubclassOptionalData()))
544 return Res;
545
546 // We have two instructions of identical opcode and #operands. Check to see
547 // if all operands are the same type
548 for (unsigned i = 0, e = L->getNumOperands(); i != e; ++i) {
549 if (int Res =
550 cmpTypes(L->getOperand(i)->getType(), R->getOperand(i)->getType()))
551 return Res;
552 }
553
554 // Check special state that is a part of some instructions.
555 if (const AllocaInst *AI = dyn_cast<AllocaInst>(L)) {
556 if (int Res = cmpTypes(AI->getAllocatedType(),
557 cast<AllocaInst>(R)->getAllocatedType()))
558 return Res;
559 return cmpNumbers(AI->getAlignment(), cast<AllocaInst>(R)->getAlignment());
560 }
561 if (const LoadInst *LI = dyn_cast<LoadInst>(L)) {
562 if (int Res = cmpNumbers(LI->isVolatile(), cast<LoadInst>(R)->isVolatile()))
563 return Res;
564 if (int Res =
565 cmpNumbers(LI->getAlignment(), cast<LoadInst>(R)->getAlignment()))
566 return Res;
567 if (int Res =
568 cmpOrderings(LI->getOrdering(), cast<LoadInst>(R)->getOrdering()))
569 return Res;
570 if (int Res = cmpNumbers(LI->getSyncScopeID(),
571 cast<LoadInst>(R)->getSyncScopeID()))
572 return Res;
573 return cmpRangeMetadata(
574 LI->getMetadata(LLVMContext::MD_range),
575 cast<LoadInst>(R)->getMetadata(LLVMContext::MD_range));
576 }
577 if (const StoreInst *SI = dyn_cast<StoreInst>(L)) {
578 if (int Res =
579 cmpNumbers(SI->isVolatile(), cast<StoreInst>(R)->isVolatile()))
580 return Res;
581 if (int Res =
582 cmpNumbers(SI->getAlignment(), cast<StoreInst>(R)->getAlignment()))
583 return Res;
584 if (int Res =
585 cmpOrderings(SI->getOrdering(), cast<StoreInst>(R)->getOrdering()))
586 return Res;
587 return cmpNumbers(SI->getSyncScopeID(),
588 cast<StoreInst>(R)->getSyncScopeID());
589 }
590 if (const CmpInst *CI = dyn_cast<CmpInst>(L))
591 return cmpNumbers(CI->getPredicate(), cast<CmpInst>(R)->getPredicate());
592 if (auto *CBL = dyn_cast<CallBase>(L)) {
593 auto *CBR = cast<CallBase>(R);
594 if (int Res = cmpNumbers(CBL->getCallingConv(), CBR->getCallingConv()))
595 return Res;
596 if (int Res = cmpAttrs(CBL->getAttributes(), CBR->getAttributes()))
597 return Res;
598 if (int Res = cmpOperandBundlesSchema(*CBL, *CBR))
599 return Res;
600 if (const CallInst *CI = dyn_cast<CallInst>(L))
601 if (int Res = cmpNumbers(CI->getTailCallKind(),
602 cast<CallInst>(R)->getTailCallKind()))
603 return Res;
604 return cmpRangeMetadata(L->getMetadata(LLVMContext::MD_range),
605 R->getMetadata(LLVMContext::MD_range));
606 }
607 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(L)) {
608 ArrayRef<unsigned> LIndices = IVI->getIndices();
609 ArrayRef<unsigned> RIndices = cast<InsertValueInst>(R)->getIndices();
610 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
611 return Res;
612 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
613 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
614 return Res;
615 }
616 return 0;
617 }
618 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(L)) {
619 ArrayRef<unsigned> LIndices = EVI->getIndices();
620 ArrayRef<unsigned> RIndices = cast<ExtractValueInst>(R)->getIndices();
621 if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
622 return Res;
623 for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
624 if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
625 return Res;
626 }
627 }
628 if (const FenceInst *FI = dyn_cast<FenceInst>(L)) {
629 if (int Res =
630 cmpOrderings(FI->getOrdering(), cast<FenceInst>(R)->getOrdering()))
631 return Res;
632 return cmpNumbers(FI->getSyncScopeID(),
633 cast<FenceInst>(R)->getSyncScopeID());
634 }
635 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(L)) {
636 if (int Res = cmpNumbers(CXI->isVolatile(),
637 cast<AtomicCmpXchgInst>(R)->isVolatile()))
638 return Res;
639 if (int Res =
640 cmpNumbers(CXI->isWeak(), cast<AtomicCmpXchgInst>(R)->isWeak()))
641 return Res;
642 if (int Res =
643 cmpOrderings(CXI->getSuccessOrdering(),
644 cast<AtomicCmpXchgInst>(R)->getSuccessOrdering()))
645 return Res;
646 if (int Res =
647 cmpOrderings(CXI->getFailureOrdering(),
648 cast<AtomicCmpXchgInst>(R)->getFailureOrdering()))
649 return Res;
650 return cmpNumbers(CXI->getSyncScopeID(),
651 cast<AtomicCmpXchgInst>(R)->getSyncScopeID());
652 }
653 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(L)) {
654 if (int Res = cmpNumbers(RMWI->getOperation(),
655 cast<AtomicRMWInst>(R)->getOperation()))
656 return Res;
657 if (int Res = cmpNumbers(RMWI->isVolatile(),
658 cast<AtomicRMWInst>(R)->isVolatile()))
659 return Res;
660 if (int Res = cmpOrderings(RMWI->getOrdering(),
661 cast<AtomicRMWInst>(R)->getOrdering()))
662 return Res;
663 return cmpNumbers(RMWI->getSyncScopeID(),
664 cast<AtomicRMWInst>(R)->getSyncScopeID());
665 }
666 if (const ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(L)) {
667 ArrayRef<int> LMask = SVI->getShuffleMask();
668 ArrayRef<int> RMask = cast<ShuffleVectorInst>(R)->getShuffleMask();
669 if (int Res = cmpNumbers(LMask.size(), RMask.size()))
670 return Res;
671 for (size_t i = 0, e = LMask.size(); i != e; ++i) {
672 if (int Res = cmpNumbers(LMask[i], RMask[i]))
673 return Res;
674 }
675 }
676 if (const PHINode *PNL = dyn_cast<PHINode>(L)) {
677 const PHINode *PNR = cast<PHINode>(R);
678 // Ensure that in addition to the incoming values being identical
679 // (checked by the caller of this function), the incoming blocks
680 // are also identical.
681 for (unsigned i = 0, e = PNL->getNumIncomingValues(); i != e; ++i) {
682 if (int Res =
683 cmpValues(PNL->getIncomingBlock(i), PNR->getIncomingBlock(i)))
684 return Res;
685 }
686 }
687 return 0;
688 }
689
690 // Determine whether two GEP operations perform the same underlying arithmetic.
691 // Read method declaration comments for more details.
cmpGEPs(const GEPOperator * GEPL,const GEPOperator * GEPR) const692 int FunctionComparator::cmpGEPs(const GEPOperator *GEPL,
693 const GEPOperator *GEPR) const {
694 unsigned int ASL = GEPL->getPointerAddressSpace();
695 unsigned int ASR = GEPR->getPointerAddressSpace();
696
697 if (int Res = cmpNumbers(ASL, ASR))
698 return Res;
699
700 // When we have target data, we can reduce the GEP down to the value in bytes
701 // added to the address.
702 const DataLayout &DL = FnL->getParent()->getDataLayout();
703 unsigned BitWidth = DL.getPointerSizeInBits(ASL);
704 APInt OffsetL(BitWidth, 0), OffsetR(BitWidth, 0);
705 if (GEPL->accumulateConstantOffset(DL, OffsetL) &&
706 GEPR->accumulateConstantOffset(DL, OffsetR))
707 return cmpAPInts(OffsetL, OffsetR);
708 if (int Res =
709 cmpTypes(GEPL->getSourceElementType(), GEPR->getSourceElementType()))
710 return Res;
711
712 if (int Res = cmpNumbers(GEPL->getNumOperands(), GEPR->getNumOperands()))
713 return Res;
714
715 for (unsigned i = 0, e = GEPL->getNumOperands(); i != e; ++i) {
716 if (int Res = cmpValues(GEPL->getOperand(i), GEPR->getOperand(i)))
717 return Res;
718 }
719
720 return 0;
721 }
722
cmpInlineAsm(const InlineAsm * L,const InlineAsm * R) const723 int FunctionComparator::cmpInlineAsm(const InlineAsm *L,
724 const InlineAsm *R) const {
725 // InlineAsm's are uniqued. If they are the same pointer, obviously they are
726 // the same, otherwise compare the fields.
727 if (L == R)
728 return 0;
729 if (int Res = cmpTypes(L->getFunctionType(), R->getFunctionType()))
730 return Res;
731 if (int Res = cmpMem(L->getAsmString(), R->getAsmString()))
732 return Res;
733 if (int Res = cmpMem(L->getConstraintString(), R->getConstraintString()))
734 return Res;
735 if (int Res = cmpNumbers(L->hasSideEffects(), R->hasSideEffects()))
736 return Res;
737 if (int Res = cmpNumbers(L->isAlignStack(), R->isAlignStack()))
738 return Res;
739 if (int Res = cmpNumbers(L->getDialect(), R->getDialect()))
740 return Res;
741 assert(L->getFunctionType() != R->getFunctionType());
742 return 0;
743 }
744
745 /// Compare two values used by the two functions under pair-wise comparison. If
746 /// this is the first time the values are seen, they're added to the mapping so
747 /// that we will detect mismatches on next use.
748 /// See comments in declaration for more details.
cmpValues(const Value * L,const Value * R) const749 int FunctionComparator::cmpValues(const Value *L, const Value *R) const {
750 // Catch self-reference case.
751 if (L == FnL) {
752 if (R == FnR)
753 return 0;
754 return -1;
755 }
756 if (R == FnR) {
757 if (L == FnL)
758 return 0;
759 return 1;
760 }
761
762 const Constant *ConstL = dyn_cast<Constant>(L);
763 const Constant *ConstR = dyn_cast<Constant>(R);
764 if (ConstL && ConstR) {
765 if (L == R)
766 return 0;
767 return cmpConstants(ConstL, ConstR);
768 }
769
770 if (ConstL)
771 return 1;
772 if (ConstR)
773 return -1;
774
775 const InlineAsm *InlineAsmL = dyn_cast<InlineAsm>(L);
776 const InlineAsm *InlineAsmR = dyn_cast<InlineAsm>(R);
777
778 if (InlineAsmL && InlineAsmR)
779 return cmpInlineAsm(InlineAsmL, InlineAsmR);
780 if (InlineAsmL)
781 return 1;
782 if (InlineAsmR)
783 return -1;
784
785 auto LeftSN = sn_mapL.insert(std::make_pair(L, sn_mapL.size())),
786 RightSN = sn_mapR.insert(std::make_pair(R, sn_mapR.size()));
787
788 return cmpNumbers(LeftSN.first->second, RightSN.first->second);
789 }
790
791 // Test whether two basic blocks have equivalent behaviour.
cmpBasicBlocks(const BasicBlock * BBL,const BasicBlock * BBR) const792 int FunctionComparator::cmpBasicBlocks(const BasicBlock *BBL,
793 const BasicBlock *BBR) const {
794 BasicBlock::const_iterator InstL = BBL->begin(), InstLE = BBL->end();
795 BasicBlock::const_iterator InstR = BBR->begin(), InstRE = BBR->end();
796
797 do {
798 bool needToCmpOperands = true;
799 if (int Res = cmpOperations(&*InstL, &*InstR, needToCmpOperands))
800 return Res;
801 if (needToCmpOperands) {
802 assert(InstL->getNumOperands() == InstR->getNumOperands());
803
804 for (unsigned i = 0, e = InstL->getNumOperands(); i != e; ++i) {
805 Value *OpL = InstL->getOperand(i);
806 Value *OpR = InstR->getOperand(i);
807 if (int Res = cmpValues(OpL, OpR))
808 return Res;
809 // cmpValues should ensure this is true.
810 assert(cmpTypes(OpL->getType(), OpR->getType()) == 0);
811 }
812 }
813
814 ++InstL;
815 ++InstR;
816 } while (InstL != InstLE && InstR != InstRE);
817
818 if (InstL != InstLE && InstR == InstRE)
819 return 1;
820 if (InstL == InstLE && InstR != InstRE)
821 return -1;
822 return 0;
823 }
824
compareSignature() const825 int FunctionComparator::compareSignature() const {
826 if (int Res = cmpAttrs(FnL->getAttributes(), FnR->getAttributes()))
827 return Res;
828
829 if (int Res = cmpNumbers(FnL->hasGC(), FnR->hasGC()))
830 return Res;
831
832 if (FnL->hasGC()) {
833 if (int Res = cmpMem(FnL->getGC(), FnR->getGC()))
834 return Res;
835 }
836
837 if (int Res = cmpNumbers(FnL->hasSection(), FnR->hasSection()))
838 return Res;
839
840 if (FnL->hasSection()) {
841 if (int Res = cmpMem(FnL->getSection(), FnR->getSection()))
842 return Res;
843 }
844
845 if (int Res = cmpNumbers(FnL->isVarArg(), FnR->isVarArg()))
846 return Res;
847
848 // TODO: if it's internal and only used in direct calls, we could handle this
849 // case too.
850 if (int Res = cmpNumbers(FnL->getCallingConv(), FnR->getCallingConv()))
851 return Res;
852
853 if (int Res = cmpTypes(FnL->getFunctionType(), FnR->getFunctionType()))
854 return Res;
855
856 assert(FnL->arg_size() == FnR->arg_size() &&
857 "Identically typed functions have different numbers of args!");
858
859 // Visit the arguments so that they get enumerated in the order they're
860 // passed in.
861 for (Function::const_arg_iterator ArgLI = FnL->arg_begin(),
862 ArgRI = FnR->arg_begin(),
863 ArgLE = FnL->arg_end();
864 ArgLI != ArgLE; ++ArgLI, ++ArgRI) {
865 if (cmpValues(&*ArgLI, &*ArgRI) != 0)
866 llvm_unreachable("Arguments repeat!");
867 }
868 return 0;
869 }
870
871 // Test whether the two functions have equivalent behaviour.
compare()872 int FunctionComparator::compare() {
873 beginCompare();
874
875 if (int Res = compareSignature())
876 return Res;
877
878 // We do a CFG-ordered walk since the actual ordering of the blocks in the
879 // linked list is immaterial. Our walk starts at the entry block for both
880 // functions, then takes each block from each terminator in order. As an
881 // artifact, this also means that unreachable blocks are ignored.
882 SmallVector<const BasicBlock *, 8> FnLBBs, FnRBBs;
883 SmallPtrSet<const BasicBlock *, 32> VisitedBBs; // in terms of F1.
884
885 FnLBBs.push_back(&FnL->getEntryBlock());
886 FnRBBs.push_back(&FnR->getEntryBlock());
887
888 VisitedBBs.insert(FnLBBs[0]);
889 while (!FnLBBs.empty()) {
890 const BasicBlock *BBL = FnLBBs.pop_back_val();
891 const BasicBlock *BBR = FnRBBs.pop_back_val();
892
893 if (int Res = cmpValues(BBL, BBR))
894 return Res;
895
896 if (int Res = cmpBasicBlocks(BBL, BBR))
897 return Res;
898
899 const Instruction *TermL = BBL->getTerminator();
900 const Instruction *TermR = BBR->getTerminator();
901
902 assert(TermL->getNumSuccessors() == TermR->getNumSuccessors());
903 for (unsigned i = 0, e = TermL->getNumSuccessors(); i != e; ++i) {
904 if (!VisitedBBs.insert(TermL->getSuccessor(i)).second)
905 continue;
906
907 FnLBBs.push_back(TermL->getSuccessor(i));
908 FnRBBs.push_back(TermR->getSuccessor(i));
909 }
910 }
911 return 0;
912 }
913
914 namespace {
915
916 // Accumulate the hash of a sequence of 64-bit integers. This is similar to a
917 // hash of a sequence of 64bit ints, but the entire input does not need to be
918 // available at once. This interface is necessary for functionHash because it
919 // needs to accumulate the hash as the structure of the function is traversed
920 // without saving these values to an intermediate buffer. This form of hashing
921 // is not often needed, as usually the object to hash is just read from a
922 // buffer.
923 class HashAccumulator64 {
924 uint64_t Hash;
925
926 public:
927 // Initialize to random constant, so the state isn't zero.
HashAccumulator64()928 HashAccumulator64() { Hash = 0x6acaa36bef8325c5ULL; }
929
add(uint64_t V)930 void add(uint64_t V) { Hash = hashing::detail::hash_16_bytes(Hash, V); }
931
932 // No finishing is required, because the entire hash value is used.
getHash()933 uint64_t getHash() { return Hash; }
934 };
935
936 } // end anonymous namespace
937
938 // A function hash is calculated by considering only the number of arguments and
939 // whether a function is varargs, the order of basic blocks (given by the
940 // successors of each basic block in depth first order), and the order of
941 // opcodes of each instruction within each of these basic blocks. This mirrors
942 // the strategy compare() uses to compare functions by walking the BBs in depth
943 // first order and comparing each instruction in sequence. Because this hash
944 // does not look at the operands, it is insensitive to things such as the
945 // target of calls and the constants used in the function, which makes it useful
946 // when possibly merging functions which are the same modulo constants and call
947 // targets.
functionHash(Function & F)948 FunctionComparator::FunctionHash FunctionComparator::functionHash(Function &F) {
949 HashAccumulator64 H;
950 H.add(F.isVarArg());
951 H.add(F.arg_size());
952
953 SmallVector<const BasicBlock *, 8> BBs;
954 SmallPtrSet<const BasicBlock *, 16> VisitedBBs;
955
956 // Walk the blocks in the same order as FunctionComparator::cmpBasicBlocks(),
957 // accumulating the hash of the function "structure." (BB and opcode sequence)
958 BBs.push_back(&F.getEntryBlock());
959 VisitedBBs.insert(BBs[0]);
960 while (!BBs.empty()) {
961 const BasicBlock *BB = BBs.pop_back_val();
962 // This random value acts as a block header, as otherwise the partition of
963 // opcodes into BBs wouldn't affect the hash, only the order of the opcodes
964 H.add(45798);
965 for (auto &Inst : *BB) {
966 H.add(Inst.getOpcode());
967 }
968 const Instruction *Term = BB->getTerminator();
969 for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) {
970 if (!VisitedBBs.insert(Term->getSuccessor(i)).second)
971 continue;
972 BBs.push_back(Term->getSuccessor(i));
973 }
974 }
975 return H.getHash();
976 }
977