1 //===------ BPFAbstractMemberAccess.cpp - Abstracting Member Accesses -----===// 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 pass abstracted struct/union member accesses in order to support 10 // compile-once run-everywhere (CO-RE). The CO-RE intends to compile the program 11 // which can run on different kernels. In particular, if bpf program tries to 12 // access a particular kernel data structure member, the details of the 13 // intermediate member access will be remembered so bpf loader can do 14 // necessary adjustment right before program loading. 15 // 16 // For example, 17 // 18 // struct s { 19 // int a; 20 // int b; 21 // }; 22 // struct t { 23 // struct s c; 24 // int d; 25 // }; 26 // struct t e; 27 // 28 // For the member access e.c.b, the compiler will generate code 29 // &e + 4 30 // 31 // The compile-once run-everywhere instead generates the following code 32 // r = 4 33 // &e + r 34 // The "4" in "r = 4" can be changed based on a particular kernel version. 35 // For example, on a particular kernel version, if struct s is changed to 36 // 37 // struct s { 38 // int new_field; 39 // int a; 40 // int b; 41 // } 42 // 43 // By repeating the member access on the host, the bpf loader can 44 // adjust "r = 4" as "r = 8". 45 // 46 // This feature relies on the following three intrinsic calls: 47 // addr = preserve_array_access_index(base, dimension, index) 48 // addr = preserve_union_access_index(base, di_index) 49 // !llvm.preserve.access.index <union_ditype> 50 // addr = preserve_struct_access_index(base, gep_index, di_index) 51 // !llvm.preserve.access.index <struct_ditype> 52 // 53 // Bitfield member access needs special attention. User cannot take the 54 // address of a bitfield acceess. To facilitate kernel verifier 55 // for easy bitfield code optimization, a new clang intrinsic is introduced: 56 // uint32_t __builtin_preserve_field_info(member_access, info_kind) 57 // In IR, a chain with two (or more) intrinsic calls will be generated: 58 // ... 59 // addr = preserve_struct_access_index(base, 1, 1) !struct s 60 // uint32_t result = bpf_preserve_field_info(addr, info_kind) 61 // 62 // Suppose the info_kind is FIELD_SIGNEDNESS, 63 // The above two IR intrinsics will be replaced with 64 // a relocatable insn: 65 // signness = /* signness of member_access */ 66 // and signness can be changed by bpf loader based on the 67 // types on the host. 68 // 69 // User can also test whether a field exists or not with 70 // uint32_t result = bpf_preserve_field_info(member_access, FIELD_EXISTENCE) 71 // The field will be always available (result = 1) during initial 72 // compilation, but bpf loader can patch with the correct value 73 // on the target host where the member_access may or may not be available 74 // 75 //===----------------------------------------------------------------------===// 76 77 #include "BPF.h" 78 #include "BPFCORE.h" 79 #include "BPFTargetMachine.h" 80 #include "llvm/BinaryFormat/Dwarf.h" 81 #include "llvm/DebugInfo/BTF/BTF.h" 82 #include "llvm/IR/DebugInfoMetadata.h" 83 #include "llvm/IR/GlobalVariable.h" 84 #include "llvm/IR/Instruction.h" 85 #include "llvm/IR/Instructions.h" 86 #include "llvm/IR/IntrinsicsBPF.h" 87 #include "llvm/IR/Module.h" 88 #include "llvm/IR/PassManager.h" 89 #include "llvm/IR/Type.h" 90 #include "llvm/IR/User.h" 91 #include "llvm/IR/Value.h" 92 #include "llvm/IR/ValueHandle.h" 93 #include "llvm/Pass.h" 94 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 95 #include <stack> 96 97 #define DEBUG_TYPE "bpf-abstract-member-access" 98 99 namespace llvm { 100 constexpr StringRef BPFCoreSharedInfo::AmaAttr; 101 uint32_t BPFCoreSharedInfo::SeqNum; 102 103 Instruction *BPFCoreSharedInfo::insertPassThrough(Module *M, BasicBlock *BB, 104 Instruction *Input, 105 Instruction *Before) { 106 Function *Fn = Intrinsic::getDeclaration( 107 M, Intrinsic::bpf_passthrough, {Input->getType(), Input->getType()}); 108 Constant *SeqNumVal = ConstantInt::get(Type::getInt32Ty(BB->getContext()), 109 BPFCoreSharedInfo::SeqNum++); 110 111 auto *NewInst = CallInst::Create(Fn, {SeqNumVal, Input}); 112 NewInst->insertBefore(Before); 113 return NewInst; 114 } 115 } // namespace llvm 116 117 using namespace llvm; 118 119 namespace { 120 class BPFAbstractMemberAccess final { 121 public: 122 BPFAbstractMemberAccess(BPFTargetMachine *TM) : TM(TM) {} 123 124 bool run(Function &F); 125 126 struct CallInfo { 127 uint32_t Kind; 128 uint32_t AccessIndex; 129 MaybeAlign RecordAlignment; 130 MDNode *Metadata; 131 WeakTrackingVH Base; 132 }; 133 typedef std::stack<std::pair<CallInst *, CallInfo>> CallInfoStack; 134 135 private: 136 enum : uint32_t { 137 BPFPreserveArrayAI = 1, 138 BPFPreserveUnionAI = 2, 139 BPFPreserveStructAI = 3, 140 BPFPreserveFieldInfoAI = 4, 141 }; 142 143 TargetMachine *TM; 144 const DataLayout *DL = nullptr; 145 Module *M = nullptr; 146 147 static std::map<std::string, GlobalVariable *> GEPGlobals; 148 // A map to link preserve_*_access_index intrinsic calls. 149 std::map<CallInst *, std::pair<CallInst *, CallInfo>> AIChain; 150 // A map to hold all the base preserve_*_access_index intrinsic calls. 151 // The base call is not an input of any other preserve_* 152 // intrinsics. 153 std::map<CallInst *, CallInfo> BaseAICalls; 154 // A map to hold <AnonRecord, TypeDef> relationships 155 std::map<DICompositeType *, DIDerivedType *> AnonRecords; 156 157 void CheckAnonRecordType(DIDerivedType *ParentTy, DIType *Ty); 158 void CheckCompositeType(DIDerivedType *ParentTy, DICompositeType *CTy); 159 void CheckDerivedType(DIDerivedType *ParentTy, DIDerivedType *DTy); 160 void ResetMetadata(struct CallInfo &CInfo); 161 162 bool doTransformation(Function &F); 163 164 void traceAICall(CallInst *Call, CallInfo &ParentInfo); 165 void traceBitCast(BitCastInst *BitCast, CallInst *Parent, 166 CallInfo &ParentInfo); 167 void traceGEP(GetElementPtrInst *GEP, CallInst *Parent, 168 CallInfo &ParentInfo); 169 void collectAICallChains(Function &F); 170 171 bool IsPreserveDIAccessIndexCall(const CallInst *Call, CallInfo &Cinfo); 172 bool IsValidAIChain(const MDNode *ParentMeta, uint32_t ParentAI, 173 const MDNode *ChildMeta); 174 bool removePreserveAccessIndexIntrinsic(Function &F); 175 bool HasPreserveFieldInfoCall(CallInfoStack &CallStack); 176 void GetStorageBitRange(DIDerivedType *MemberTy, Align RecordAlignment, 177 uint32_t &StartBitOffset, uint32_t &EndBitOffset); 178 uint32_t GetFieldInfo(uint32_t InfoKind, DICompositeType *CTy, 179 uint32_t AccessIndex, uint32_t PatchImm, 180 MaybeAlign RecordAlignment); 181 182 Value *computeBaseAndAccessKey(CallInst *Call, CallInfo &CInfo, 183 std::string &AccessKey, MDNode *&BaseMeta); 184 MDNode *computeAccessKey(CallInst *Call, CallInfo &CInfo, 185 std::string &AccessKey, bool &IsInt32Ret); 186 bool transformGEPChain(CallInst *Call, CallInfo &CInfo); 187 }; 188 189 std::map<std::string, GlobalVariable *> BPFAbstractMemberAccess::GEPGlobals; 190 } // End anonymous namespace 191 192 bool BPFAbstractMemberAccess::run(Function &F) { 193 LLVM_DEBUG(dbgs() << "********** Abstract Member Accesses **********\n"); 194 195 M = F.getParent(); 196 if (!M) 197 return false; 198 199 // Bail out if no debug info. 200 if (M->debug_compile_units().empty()) 201 return false; 202 203 // For each argument/return/local_variable type, trace the type 204 // pattern like '[derived_type]* [composite_type]' to check 205 // and remember (anon record -> typedef) relations where the 206 // anon record is defined as 207 // typedef [const/volatile/restrict]* [anon record] 208 DISubprogram *SP = F.getSubprogram(); 209 if (SP && SP->isDefinition()) { 210 for (DIType *Ty: SP->getType()->getTypeArray()) 211 CheckAnonRecordType(nullptr, Ty); 212 for (const DINode *DN : SP->getRetainedNodes()) { 213 if (const auto *DV = dyn_cast<DILocalVariable>(DN)) 214 CheckAnonRecordType(nullptr, DV->getType()); 215 } 216 } 217 218 DL = &M->getDataLayout(); 219 return doTransformation(F); 220 } 221 222 void BPFAbstractMemberAccess::ResetMetadata(struct CallInfo &CInfo) { 223 if (auto Ty = dyn_cast<DICompositeType>(CInfo.Metadata)) { 224 if (AnonRecords.find(Ty) != AnonRecords.end()) { 225 if (AnonRecords[Ty] != nullptr) 226 CInfo.Metadata = AnonRecords[Ty]; 227 } 228 } 229 } 230 231 void BPFAbstractMemberAccess::CheckCompositeType(DIDerivedType *ParentTy, 232 DICompositeType *CTy) { 233 if (!CTy->getName().empty() || !ParentTy || 234 ParentTy->getTag() != dwarf::DW_TAG_typedef) 235 return; 236 237 if (AnonRecords.find(CTy) == AnonRecords.end()) { 238 AnonRecords[CTy] = ParentTy; 239 return; 240 } 241 242 // Two or more typedef's may point to the same anon record. 243 // If this is the case, set the typedef DIType to be nullptr 244 // to indicate the duplication case. 245 DIDerivedType *CurrTy = AnonRecords[CTy]; 246 if (CurrTy == ParentTy) 247 return; 248 AnonRecords[CTy] = nullptr; 249 } 250 251 void BPFAbstractMemberAccess::CheckDerivedType(DIDerivedType *ParentTy, 252 DIDerivedType *DTy) { 253 DIType *BaseType = DTy->getBaseType(); 254 if (!BaseType) 255 return; 256 257 unsigned Tag = DTy->getTag(); 258 if (Tag == dwarf::DW_TAG_pointer_type) 259 CheckAnonRecordType(nullptr, BaseType); 260 else if (Tag == dwarf::DW_TAG_typedef) 261 CheckAnonRecordType(DTy, BaseType); 262 else 263 CheckAnonRecordType(ParentTy, BaseType); 264 } 265 266 void BPFAbstractMemberAccess::CheckAnonRecordType(DIDerivedType *ParentTy, 267 DIType *Ty) { 268 if (!Ty) 269 return; 270 271 if (auto *CTy = dyn_cast<DICompositeType>(Ty)) 272 return CheckCompositeType(ParentTy, CTy); 273 else if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) 274 return CheckDerivedType(ParentTy, DTy); 275 } 276 277 static bool SkipDIDerivedTag(unsigned Tag, bool skipTypedef) { 278 if (Tag != dwarf::DW_TAG_typedef && Tag != dwarf::DW_TAG_const_type && 279 Tag != dwarf::DW_TAG_volatile_type && 280 Tag != dwarf::DW_TAG_restrict_type && 281 Tag != dwarf::DW_TAG_member) 282 return false; 283 if (Tag == dwarf::DW_TAG_typedef && !skipTypedef) 284 return false; 285 return true; 286 } 287 288 static DIType * stripQualifiers(DIType *Ty, bool skipTypedef = true) { 289 while (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 290 if (!SkipDIDerivedTag(DTy->getTag(), skipTypedef)) 291 break; 292 Ty = DTy->getBaseType(); 293 } 294 return Ty; 295 } 296 297 static const DIType * stripQualifiers(const DIType *Ty) { 298 while (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 299 if (!SkipDIDerivedTag(DTy->getTag(), true)) 300 break; 301 Ty = DTy->getBaseType(); 302 } 303 return Ty; 304 } 305 306 static uint32_t calcArraySize(const DICompositeType *CTy, uint32_t StartDim) { 307 DINodeArray Elements = CTy->getElements(); 308 uint32_t DimSize = 1; 309 for (uint32_t I = StartDim; I < Elements.size(); ++I) { 310 if (auto *Element = dyn_cast_or_null<DINode>(Elements[I])) 311 if (Element->getTag() == dwarf::DW_TAG_subrange_type) { 312 const DISubrange *SR = cast<DISubrange>(Element); 313 auto *CI = SR->getCount().dyn_cast<ConstantInt *>(); 314 DimSize *= CI->getSExtValue(); 315 } 316 } 317 318 return DimSize; 319 } 320 321 static Type *getBaseElementType(const CallInst *Call) { 322 // Element type is stored in an elementtype() attribute on the first param. 323 return Call->getParamElementType(0); 324 } 325 326 static uint64_t getConstant(const Value *IndexValue) { 327 const ConstantInt *CV = dyn_cast<ConstantInt>(IndexValue); 328 assert(CV); 329 return CV->getValue().getZExtValue(); 330 } 331 332 /// Check whether a call is a preserve_*_access_index intrinsic call or not. 333 bool BPFAbstractMemberAccess::IsPreserveDIAccessIndexCall(const CallInst *Call, 334 CallInfo &CInfo) { 335 if (!Call) 336 return false; 337 338 const auto *GV = dyn_cast<GlobalValue>(Call->getCalledOperand()); 339 if (!GV) 340 return false; 341 if (GV->getName().starts_with("llvm.preserve.array.access.index")) { 342 CInfo.Kind = BPFPreserveArrayAI; 343 CInfo.Metadata = Call->getMetadata(LLVMContext::MD_preserve_access_index); 344 if (!CInfo.Metadata) 345 report_fatal_error("Missing metadata for llvm.preserve.array.access.index intrinsic"); 346 CInfo.AccessIndex = getConstant(Call->getArgOperand(2)); 347 CInfo.Base = Call->getArgOperand(0); 348 CInfo.RecordAlignment = DL->getABITypeAlign(getBaseElementType(Call)); 349 return true; 350 } 351 if (GV->getName().starts_with("llvm.preserve.union.access.index")) { 352 CInfo.Kind = BPFPreserveUnionAI; 353 CInfo.Metadata = Call->getMetadata(LLVMContext::MD_preserve_access_index); 354 if (!CInfo.Metadata) 355 report_fatal_error("Missing metadata for llvm.preserve.union.access.index intrinsic"); 356 ResetMetadata(CInfo); 357 CInfo.AccessIndex = getConstant(Call->getArgOperand(1)); 358 CInfo.Base = Call->getArgOperand(0); 359 return true; 360 } 361 if (GV->getName().starts_with("llvm.preserve.struct.access.index")) { 362 CInfo.Kind = BPFPreserveStructAI; 363 CInfo.Metadata = Call->getMetadata(LLVMContext::MD_preserve_access_index); 364 if (!CInfo.Metadata) 365 report_fatal_error("Missing metadata for llvm.preserve.struct.access.index intrinsic"); 366 ResetMetadata(CInfo); 367 CInfo.AccessIndex = getConstant(Call->getArgOperand(2)); 368 CInfo.Base = Call->getArgOperand(0); 369 CInfo.RecordAlignment = DL->getABITypeAlign(getBaseElementType(Call)); 370 return true; 371 } 372 if (GV->getName().starts_with("llvm.bpf.preserve.field.info")) { 373 CInfo.Kind = BPFPreserveFieldInfoAI; 374 CInfo.Metadata = nullptr; 375 // Check validity of info_kind as clang did not check this. 376 uint64_t InfoKind = getConstant(Call->getArgOperand(1)); 377 if (InfoKind >= BTF::MAX_FIELD_RELOC_KIND) 378 report_fatal_error("Incorrect info_kind for llvm.bpf.preserve.field.info intrinsic"); 379 CInfo.AccessIndex = InfoKind; 380 return true; 381 } 382 if (GV->getName().starts_with("llvm.bpf.preserve.type.info")) { 383 CInfo.Kind = BPFPreserveFieldInfoAI; 384 CInfo.Metadata = Call->getMetadata(LLVMContext::MD_preserve_access_index); 385 if (!CInfo.Metadata) 386 report_fatal_error("Missing metadata for llvm.preserve.type.info intrinsic"); 387 uint64_t Flag = getConstant(Call->getArgOperand(1)); 388 if (Flag >= BPFCoreSharedInfo::MAX_PRESERVE_TYPE_INFO_FLAG) 389 report_fatal_error("Incorrect flag for llvm.bpf.preserve.type.info intrinsic"); 390 if (Flag == BPFCoreSharedInfo::PRESERVE_TYPE_INFO_EXISTENCE) 391 CInfo.AccessIndex = BTF::TYPE_EXISTENCE; 392 else if (Flag == BPFCoreSharedInfo::PRESERVE_TYPE_INFO_MATCH) 393 CInfo.AccessIndex = BTF::TYPE_MATCH; 394 else 395 CInfo.AccessIndex = BTF::TYPE_SIZE; 396 return true; 397 } 398 if (GV->getName().starts_with("llvm.bpf.preserve.enum.value")) { 399 CInfo.Kind = BPFPreserveFieldInfoAI; 400 CInfo.Metadata = Call->getMetadata(LLVMContext::MD_preserve_access_index); 401 if (!CInfo.Metadata) 402 report_fatal_error("Missing metadata for llvm.preserve.enum.value intrinsic"); 403 uint64_t Flag = getConstant(Call->getArgOperand(2)); 404 if (Flag >= BPFCoreSharedInfo::MAX_PRESERVE_ENUM_VALUE_FLAG) 405 report_fatal_error("Incorrect flag for llvm.bpf.preserve.enum.value intrinsic"); 406 if (Flag == BPFCoreSharedInfo::PRESERVE_ENUM_VALUE_EXISTENCE) 407 CInfo.AccessIndex = BTF::ENUM_VALUE_EXISTENCE; 408 else 409 CInfo.AccessIndex = BTF::ENUM_VALUE; 410 return true; 411 } 412 413 return false; 414 } 415 416 static void replaceWithGEP(CallInst *Call, uint32_t DimensionIndex, 417 uint32_t GEPIndex) { 418 uint32_t Dimension = 1; 419 if (DimensionIndex > 0) 420 Dimension = getConstant(Call->getArgOperand(DimensionIndex)); 421 422 Constant *Zero = 423 ConstantInt::get(Type::getInt32Ty(Call->getParent()->getContext()), 0); 424 SmallVector<Value *, 4> IdxList; 425 for (unsigned I = 0; I < Dimension; ++I) 426 IdxList.push_back(Zero); 427 IdxList.push_back(Call->getArgOperand(GEPIndex)); 428 429 auto *GEP = GetElementPtrInst::CreateInBounds( 430 getBaseElementType(Call), Call->getArgOperand(0), IdxList, "", Call); 431 Call->replaceAllUsesWith(GEP); 432 Call->eraseFromParent(); 433 } 434 435 void BPFCoreSharedInfo::removeArrayAccessCall(CallInst *Call) { 436 replaceWithGEP(Call, 1, 2); 437 } 438 439 void BPFCoreSharedInfo::removeStructAccessCall(CallInst *Call) { 440 replaceWithGEP(Call, 0, 1); 441 } 442 443 void BPFCoreSharedInfo::removeUnionAccessCall(CallInst *Call) { 444 Call->replaceAllUsesWith(Call->getArgOperand(0)); 445 Call->eraseFromParent(); 446 } 447 448 bool BPFAbstractMemberAccess::removePreserveAccessIndexIntrinsic(Function &F) { 449 std::vector<CallInst *> PreserveArrayIndexCalls; 450 std::vector<CallInst *> PreserveUnionIndexCalls; 451 std::vector<CallInst *> PreserveStructIndexCalls; 452 bool Found = false; 453 454 for (auto &BB : F) 455 for (auto &I : BB) { 456 auto *Call = dyn_cast<CallInst>(&I); 457 CallInfo CInfo; 458 if (!IsPreserveDIAccessIndexCall(Call, CInfo)) 459 continue; 460 461 Found = true; 462 if (CInfo.Kind == BPFPreserveArrayAI) 463 PreserveArrayIndexCalls.push_back(Call); 464 else if (CInfo.Kind == BPFPreserveUnionAI) 465 PreserveUnionIndexCalls.push_back(Call); 466 else 467 PreserveStructIndexCalls.push_back(Call); 468 } 469 470 // do the following transformation: 471 // . addr = preserve_array_access_index(base, dimension, index) 472 // is transformed to 473 // addr = GEP(base, dimenion's zero's, index) 474 // . addr = preserve_union_access_index(base, di_index) 475 // is transformed to 476 // addr = base, i.e., all usages of "addr" are replaced by "base". 477 // . addr = preserve_struct_access_index(base, gep_index, di_index) 478 // is transformed to 479 // addr = GEP(base, 0, gep_index) 480 for (CallInst *Call : PreserveArrayIndexCalls) 481 BPFCoreSharedInfo::removeArrayAccessCall(Call); 482 for (CallInst *Call : PreserveStructIndexCalls) 483 BPFCoreSharedInfo::removeStructAccessCall(Call); 484 for (CallInst *Call : PreserveUnionIndexCalls) 485 BPFCoreSharedInfo::removeUnionAccessCall(Call); 486 487 return Found; 488 } 489 490 /// Check whether the access index chain is valid. We check 491 /// here because there may be type casts between two 492 /// access indexes. We want to ensure memory access still valid. 493 bool BPFAbstractMemberAccess::IsValidAIChain(const MDNode *ParentType, 494 uint32_t ParentAI, 495 const MDNode *ChildType) { 496 if (!ChildType) 497 return true; // preserve_field_info, no type comparison needed. 498 499 const DIType *PType = stripQualifiers(cast<DIType>(ParentType)); 500 const DIType *CType = stripQualifiers(cast<DIType>(ChildType)); 501 502 // Child is a derived/pointer type, which is due to type casting. 503 // Pointer type cannot be in the middle of chain. 504 if (isa<DIDerivedType>(CType)) 505 return false; 506 507 // Parent is a pointer type. 508 if (const auto *PtrTy = dyn_cast<DIDerivedType>(PType)) { 509 if (PtrTy->getTag() != dwarf::DW_TAG_pointer_type) 510 return false; 511 return stripQualifiers(PtrTy->getBaseType()) == CType; 512 } 513 514 // Otherwise, struct/union/array types 515 const auto *PTy = dyn_cast<DICompositeType>(PType); 516 const auto *CTy = dyn_cast<DICompositeType>(CType); 517 assert(PTy && CTy && "ParentType or ChildType is null or not composite"); 518 519 uint32_t PTyTag = PTy->getTag(); 520 assert(PTyTag == dwarf::DW_TAG_array_type || 521 PTyTag == dwarf::DW_TAG_structure_type || 522 PTyTag == dwarf::DW_TAG_union_type); 523 524 uint32_t CTyTag = CTy->getTag(); 525 assert(CTyTag == dwarf::DW_TAG_array_type || 526 CTyTag == dwarf::DW_TAG_structure_type || 527 CTyTag == dwarf::DW_TAG_union_type); 528 529 // Multi dimensional arrays, base element should be the same 530 if (PTyTag == dwarf::DW_TAG_array_type && PTyTag == CTyTag) 531 return PTy->getBaseType() == CTy->getBaseType(); 532 533 DIType *Ty; 534 if (PTyTag == dwarf::DW_TAG_array_type) 535 Ty = PTy->getBaseType(); 536 else 537 Ty = dyn_cast<DIType>(PTy->getElements()[ParentAI]); 538 539 return dyn_cast<DICompositeType>(stripQualifiers(Ty)) == CTy; 540 } 541 542 void BPFAbstractMemberAccess::traceAICall(CallInst *Call, 543 CallInfo &ParentInfo) { 544 for (User *U : Call->users()) { 545 Instruction *Inst = dyn_cast<Instruction>(U); 546 if (!Inst) 547 continue; 548 549 if (auto *BI = dyn_cast<BitCastInst>(Inst)) { 550 traceBitCast(BI, Call, ParentInfo); 551 } else if (auto *CI = dyn_cast<CallInst>(Inst)) { 552 CallInfo ChildInfo; 553 554 if (IsPreserveDIAccessIndexCall(CI, ChildInfo) && 555 IsValidAIChain(ParentInfo.Metadata, ParentInfo.AccessIndex, 556 ChildInfo.Metadata)) { 557 AIChain[CI] = std::make_pair(Call, ParentInfo); 558 traceAICall(CI, ChildInfo); 559 } else { 560 BaseAICalls[Call] = ParentInfo; 561 } 562 } else if (auto *GI = dyn_cast<GetElementPtrInst>(Inst)) { 563 if (GI->hasAllZeroIndices()) 564 traceGEP(GI, Call, ParentInfo); 565 else 566 BaseAICalls[Call] = ParentInfo; 567 } else { 568 BaseAICalls[Call] = ParentInfo; 569 } 570 } 571 } 572 573 void BPFAbstractMemberAccess::traceBitCast(BitCastInst *BitCast, 574 CallInst *Parent, 575 CallInfo &ParentInfo) { 576 for (User *U : BitCast->users()) { 577 Instruction *Inst = dyn_cast<Instruction>(U); 578 if (!Inst) 579 continue; 580 581 if (auto *BI = dyn_cast<BitCastInst>(Inst)) { 582 traceBitCast(BI, Parent, ParentInfo); 583 } else if (auto *CI = dyn_cast<CallInst>(Inst)) { 584 CallInfo ChildInfo; 585 if (IsPreserveDIAccessIndexCall(CI, ChildInfo) && 586 IsValidAIChain(ParentInfo.Metadata, ParentInfo.AccessIndex, 587 ChildInfo.Metadata)) { 588 AIChain[CI] = std::make_pair(Parent, ParentInfo); 589 traceAICall(CI, ChildInfo); 590 } else { 591 BaseAICalls[Parent] = ParentInfo; 592 } 593 } else if (auto *GI = dyn_cast<GetElementPtrInst>(Inst)) { 594 if (GI->hasAllZeroIndices()) 595 traceGEP(GI, Parent, ParentInfo); 596 else 597 BaseAICalls[Parent] = ParentInfo; 598 } else { 599 BaseAICalls[Parent] = ParentInfo; 600 } 601 } 602 } 603 604 void BPFAbstractMemberAccess::traceGEP(GetElementPtrInst *GEP, CallInst *Parent, 605 CallInfo &ParentInfo) { 606 for (User *U : GEP->users()) { 607 Instruction *Inst = dyn_cast<Instruction>(U); 608 if (!Inst) 609 continue; 610 611 if (auto *BI = dyn_cast<BitCastInst>(Inst)) { 612 traceBitCast(BI, Parent, ParentInfo); 613 } else if (auto *CI = dyn_cast<CallInst>(Inst)) { 614 CallInfo ChildInfo; 615 if (IsPreserveDIAccessIndexCall(CI, ChildInfo) && 616 IsValidAIChain(ParentInfo.Metadata, ParentInfo.AccessIndex, 617 ChildInfo.Metadata)) { 618 AIChain[CI] = std::make_pair(Parent, ParentInfo); 619 traceAICall(CI, ChildInfo); 620 } else { 621 BaseAICalls[Parent] = ParentInfo; 622 } 623 } else if (auto *GI = dyn_cast<GetElementPtrInst>(Inst)) { 624 if (GI->hasAllZeroIndices()) 625 traceGEP(GI, Parent, ParentInfo); 626 else 627 BaseAICalls[Parent] = ParentInfo; 628 } else { 629 BaseAICalls[Parent] = ParentInfo; 630 } 631 } 632 } 633 634 void BPFAbstractMemberAccess::collectAICallChains(Function &F) { 635 AIChain.clear(); 636 BaseAICalls.clear(); 637 638 for (auto &BB : F) 639 for (auto &I : BB) { 640 CallInfo CInfo; 641 auto *Call = dyn_cast<CallInst>(&I); 642 if (!IsPreserveDIAccessIndexCall(Call, CInfo) || 643 AIChain.find(Call) != AIChain.end()) 644 continue; 645 646 traceAICall(Call, CInfo); 647 } 648 } 649 650 /// Get the start and the end of storage offset for \p MemberTy. 651 void BPFAbstractMemberAccess::GetStorageBitRange(DIDerivedType *MemberTy, 652 Align RecordAlignment, 653 uint32_t &StartBitOffset, 654 uint32_t &EndBitOffset) { 655 uint32_t MemberBitSize = MemberTy->getSizeInBits(); 656 uint32_t MemberBitOffset = MemberTy->getOffsetInBits(); 657 658 if (RecordAlignment > 8) { 659 // If the Bits are within an aligned 8-byte, set the RecordAlignment 660 // to 8, other report the fatal error. 661 if (MemberBitOffset / 64 != (MemberBitOffset + MemberBitSize) / 64) 662 report_fatal_error("Unsupported field expression for llvm.bpf.preserve.field.info, " 663 "requiring too big alignment"); 664 RecordAlignment = Align(8); 665 } 666 667 uint32_t AlignBits = RecordAlignment.value() * 8; 668 if (MemberBitSize > AlignBits) 669 report_fatal_error("Unsupported field expression for llvm.bpf.preserve.field.info, " 670 "bitfield size greater than record alignment"); 671 672 StartBitOffset = MemberBitOffset & ~(AlignBits - 1); 673 if ((StartBitOffset + AlignBits) < (MemberBitOffset + MemberBitSize)) 674 report_fatal_error("Unsupported field expression for llvm.bpf.preserve.field.info, " 675 "cross alignment boundary"); 676 EndBitOffset = StartBitOffset + AlignBits; 677 } 678 679 uint32_t BPFAbstractMemberAccess::GetFieldInfo(uint32_t InfoKind, 680 DICompositeType *CTy, 681 uint32_t AccessIndex, 682 uint32_t PatchImm, 683 MaybeAlign RecordAlignment) { 684 if (InfoKind == BTF::FIELD_EXISTENCE) 685 return 1; 686 687 uint32_t Tag = CTy->getTag(); 688 if (InfoKind == BTF::FIELD_BYTE_OFFSET) { 689 if (Tag == dwarf::DW_TAG_array_type) { 690 auto *EltTy = stripQualifiers(CTy->getBaseType()); 691 PatchImm += AccessIndex * calcArraySize(CTy, 1) * 692 (EltTy->getSizeInBits() >> 3); 693 } else if (Tag == dwarf::DW_TAG_structure_type) { 694 auto *MemberTy = cast<DIDerivedType>(CTy->getElements()[AccessIndex]); 695 if (!MemberTy->isBitField()) { 696 PatchImm += MemberTy->getOffsetInBits() >> 3; 697 } else { 698 unsigned SBitOffset, NextSBitOffset; 699 GetStorageBitRange(MemberTy, *RecordAlignment, SBitOffset, 700 NextSBitOffset); 701 PatchImm += SBitOffset >> 3; 702 } 703 } 704 return PatchImm; 705 } 706 707 if (InfoKind == BTF::FIELD_BYTE_SIZE) { 708 if (Tag == dwarf::DW_TAG_array_type) { 709 auto *EltTy = stripQualifiers(CTy->getBaseType()); 710 return calcArraySize(CTy, 1) * (EltTy->getSizeInBits() >> 3); 711 } else { 712 auto *MemberTy = cast<DIDerivedType>(CTy->getElements()[AccessIndex]); 713 uint32_t SizeInBits = MemberTy->getSizeInBits(); 714 if (!MemberTy->isBitField()) 715 return SizeInBits >> 3; 716 717 unsigned SBitOffset, NextSBitOffset; 718 GetStorageBitRange(MemberTy, *RecordAlignment, SBitOffset, 719 NextSBitOffset); 720 SizeInBits = NextSBitOffset - SBitOffset; 721 if (SizeInBits & (SizeInBits - 1)) 722 report_fatal_error("Unsupported field expression for llvm.bpf.preserve.field.info"); 723 return SizeInBits >> 3; 724 } 725 } 726 727 if (InfoKind == BTF::FIELD_SIGNEDNESS) { 728 const DIType *BaseTy; 729 if (Tag == dwarf::DW_TAG_array_type) { 730 // Signedness only checked when final array elements are accessed. 731 if (CTy->getElements().size() != 1) 732 report_fatal_error("Invalid array expression for llvm.bpf.preserve.field.info"); 733 BaseTy = stripQualifiers(CTy->getBaseType()); 734 } else { 735 auto *MemberTy = cast<DIDerivedType>(CTy->getElements()[AccessIndex]); 736 BaseTy = stripQualifiers(MemberTy->getBaseType()); 737 } 738 739 // Only basic types and enum types have signedness. 740 const auto *BTy = dyn_cast<DIBasicType>(BaseTy); 741 while (!BTy) { 742 const auto *CompTy = dyn_cast<DICompositeType>(BaseTy); 743 // Report an error if the field expression does not have signedness. 744 if (!CompTy || CompTy->getTag() != dwarf::DW_TAG_enumeration_type) 745 report_fatal_error("Invalid field expression for llvm.bpf.preserve.field.info"); 746 BaseTy = stripQualifiers(CompTy->getBaseType()); 747 BTy = dyn_cast<DIBasicType>(BaseTy); 748 } 749 uint32_t Encoding = BTy->getEncoding(); 750 return (Encoding == dwarf::DW_ATE_signed || Encoding == dwarf::DW_ATE_signed_char); 751 } 752 753 if (InfoKind == BTF::FIELD_LSHIFT_U64) { 754 // The value is loaded into a value with FIELD_BYTE_SIZE size, 755 // and then zero or sign extended to U64. 756 // FIELD_LSHIFT_U64 and FIELD_RSHIFT_U64 are operations 757 // to extract the original value. 758 const Triple &Triple = TM->getTargetTriple(); 759 DIDerivedType *MemberTy = nullptr; 760 bool IsBitField = false; 761 uint32_t SizeInBits; 762 763 if (Tag == dwarf::DW_TAG_array_type) { 764 auto *EltTy = stripQualifiers(CTy->getBaseType()); 765 SizeInBits = calcArraySize(CTy, 1) * EltTy->getSizeInBits(); 766 } else { 767 MemberTy = cast<DIDerivedType>(CTy->getElements()[AccessIndex]); 768 SizeInBits = MemberTy->getSizeInBits(); 769 IsBitField = MemberTy->isBitField(); 770 } 771 772 if (!IsBitField) { 773 if (SizeInBits > 64) 774 report_fatal_error("too big field size for llvm.bpf.preserve.field.info"); 775 return 64 - SizeInBits; 776 } 777 778 unsigned SBitOffset, NextSBitOffset; 779 GetStorageBitRange(MemberTy, *RecordAlignment, SBitOffset, NextSBitOffset); 780 if (NextSBitOffset - SBitOffset > 64) 781 report_fatal_error("too big field size for llvm.bpf.preserve.field.info"); 782 783 unsigned OffsetInBits = MemberTy->getOffsetInBits(); 784 if (Triple.getArch() == Triple::bpfel) 785 return SBitOffset + 64 - OffsetInBits - SizeInBits; 786 else 787 return OffsetInBits + 64 - NextSBitOffset; 788 } 789 790 if (InfoKind == BTF::FIELD_RSHIFT_U64) { 791 DIDerivedType *MemberTy = nullptr; 792 bool IsBitField = false; 793 uint32_t SizeInBits; 794 if (Tag == dwarf::DW_TAG_array_type) { 795 auto *EltTy = stripQualifiers(CTy->getBaseType()); 796 SizeInBits = calcArraySize(CTy, 1) * EltTy->getSizeInBits(); 797 } else { 798 MemberTy = cast<DIDerivedType>(CTy->getElements()[AccessIndex]); 799 SizeInBits = MemberTy->getSizeInBits(); 800 IsBitField = MemberTy->isBitField(); 801 } 802 803 if (!IsBitField) { 804 if (SizeInBits > 64) 805 report_fatal_error("too big field size for llvm.bpf.preserve.field.info"); 806 return 64 - SizeInBits; 807 } 808 809 unsigned SBitOffset, NextSBitOffset; 810 GetStorageBitRange(MemberTy, *RecordAlignment, SBitOffset, NextSBitOffset); 811 if (NextSBitOffset - SBitOffset > 64) 812 report_fatal_error("too big field size for llvm.bpf.preserve.field.info"); 813 814 return 64 - SizeInBits; 815 } 816 817 llvm_unreachable("Unknown llvm.bpf.preserve.field.info info kind"); 818 } 819 820 bool BPFAbstractMemberAccess::HasPreserveFieldInfoCall(CallInfoStack &CallStack) { 821 // This is called in error return path, no need to maintain CallStack. 822 while (CallStack.size()) { 823 auto StackElem = CallStack.top(); 824 if (StackElem.second.Kind == BPFPreserveFieldInfoAI) 825 return true; 826 CallStack.pop(); 827 } 828 return false; 829 } 830 831 /// Compute the base of the whole preserve_* intrinsics chains, i.e., the base 832 /// pointer of the first preserve_*_access_index call, and construct the access 833 /// string, which will be the name of a global variable. 834 Value *BPFAbstractMemberAccess::computeBaseAndAccessKey(CallInst *Call, 835 CallInfo &CInfo, 836 std::string &AccessKey, 837 MDNode *&TypeMeta) { 838 Value *Base = nullptr; 839 std::string TypeName; 840 CallInfoStack CallStack; 841 842 // Put the access chain into a stack with the top as the head of the chain. 843 while (Call) { 844 CallStack.push(std::make_pair(Call, CInfo)); 845 CInfo = AIChain[Call].second; 846 Call = AIChain[Call].first; 847 } 848 849 // The access offset from the base of the head of chain is also 850 // calculated here as all debuginfo types are available. 851 852 // Get type name and calculate the first index. 853 // We only want to get type name from typedef, structure or union. 854 // If user wants a relocation like 855 // int *p; ... __builtin_preserve_access_index(&p[4]) ... 856 // or 857 // int a[10][20]; ... __builtin_preserve_access_index(&a[2][3]) ... 858 // we will skip them. 859 uint32_t FirstIndex = 0; 860 uint32_t PatchImm = 0; // AccessOffset or the requested field info 861 uint32_t InfoKind = BTF::FIELD_BYTE_OFFSET; 862 while (CallStack.size()) { 863 auto StackElem = CallStack.top(); 864 Call = StackElem.first; 865 CInfo = StackElem.second; 866 867 if (!Base) 868 Base = CInfo.Base; 869 870 DIType *PossibleTypeDef = stripQualifiers(cast<DIType>(CInfo.Metadata), 871 false); 872 DIType *Ty = stripQualifiers(PossibleTypeDef); 873 if (CInfo.Kind == BPFPreserveUnionAI || 874 CInfo.Kind == BPFPreserveStructAI) { 875 // struct or union type. If the typedef is in the metadata, always 876 // use the typedef. 877 TypeName = std::string(PossibleTypeDef->getName()); 878 TypeMeta = PossibleTypeDef; 879 PatchImm += FirstIndex * (Ty->getSizeInBits() >> 3); 880 break; 881 } 882 883 assert(CInfo.Kind == BPFPreserveArrayAI); 884 885 // Array entries will always be consumed for accumulative initial index. 886 CallStack.pop(); 887 888 // BPFPreserveArrayAI 889 uint64_t AccessIndex = CInfo.AccessIndex; 890 891 DIType *BaseTy = nullptr; 892 bool CheckElemType = false; 893 if (const auto *CTy = dyn_cast<DICompositeType>(Ty)) { 894 // array type 895 assert(CTy->getTag() == dwarf::DW_TAG_array_type); 896 897 898 FirstIndex += AccessIndex * calcArraySize(CTy, 1); 899 BaseTy = stripQualifiers(CTy->getBaseType()); 900 CheckElemType = CTy->getElements().size() == 1; 901 } else { 902 // pointer type 903 auto *DTy = cast<DIDerivedType>(Ty); 904 assert(DTy->getTag() == dwarf::DW_TAG_pointer_type); 905 906 BaseTy = stripQualifiers(DTy->getBaseType()); 907 CTy = dyn_cast<DICompositeType>(BaseTy); 908 if (!CTy) { 909 CheckElemType = true; 910 } else if (CTy->getTag() != dwarf::DW_TAG_array_type) { 911 FirstIndex += AccessIndex; 912 CheckElemType = true; 913 } else { 914 FirstIndex += AccessIndex * calcArraySize(CTy, 0); 915 } 916 } 917 918 if (CheckElemType) { 919 auto *CTy = dyn_cast<DICompositeType>(BaseTy); 920 if (!CTy) { 921 if (HasPreserveFieldInfoCall(CallStack)) 922 report_fatal_error("Invalid field access for llvm.preserve.field.info intrinsic"); 923 return nullptr; 924 } 925 926 unsigned CTag = CTy->getTag(); 927 if (CTag == dwarf::DW_TAG_structure_type || CTag == dwarf::DW_TAG_union_type) { 928 TypeName = std::string(CTy->getName()); 929 } else { 930 if (HasPreserveFieldInfoCall(CallStack)) 931 report_fatal_error("Invalid field access for llvm.preserve.field.info intrinsic"); 932 return nullptr; 933 } 934 TypeMeta = CTy; 935 PatchImm += FirstIndex * (CTy->getSizeInBits() >> 3); 936 break; 937 } 938 } 939 assert(TypeName.size()); 940 AccessKey += std::to_string(FirstIndex); 941 942 // Traverse the rest of access chain to complete offset calculation 943 // and access key construction. 944 while (CallStack.size()) { 945 auto StackElem = CallStack.top(); 946 CInfo = StackElem.second; 947 CallStack.pop(); 948 949 if (CInfo.Kind == BPFPreserveFieldInfoAI) { 950 InfoKind = CInfo.AccessIndex; 951 if (InfoKind == BTF::FIELD_EXISTENCE) 952 PatchImm = 1; 953 break; 954 } 955 956 // If the next Call (the top of the stack) is a BPFPreserveFieldInfoAI, 957 // the action will be extracting field info. 958 if (CallStack.size()) { 959 auto StackElem2 = CallStack.top(); 960 CallInfo CInfo2 = StackElem2.second; 961 if (CInfo2.Kind == BPFPreserveFieldInfoAI) { 962 InfoKind = CInfo2.AccessIndex; 963 assert(CallStack.size() == 1); 964 } 965 } 966 967 // Access Index 968 uint64_t AccessIndex = CInfo.AccessIndex; 969 AccessKey += ":" + std::to_string(AccessIndex); 970 971 MDNode *MDN = CInfo.Metadata; 972 // At this stage, it cannot be pointer type. 973 auto *CTy = cast<DICompositeType>(stripQualifiers(cast<DIType>(MDN))); 974 PatchImm = GetFieldInfo(InfoKind, CTy, AccessIndex, PatchImm, 975 CInfo.RecordAlignment); 976 } 977 978 // Access key is the 979 // "llvm." + type name + ":" + reloc type + ":" + patched imm + "$" + 980 // access string, 981 // uniquely identifying one relocation. 982 // The prefix "llvm." indicates this is a temporary global, which should 983 // not be emitted to ELF file. 984 AccessKey = "llvm." + TypeName + ":" + std::to_string(InfoKind) + ":" + 985 std::to_string(PatchImm) + "$" + AccessKey; 986 987 return Base; 988 } 989 990 MDNode *BPFAbstractMemberAccess::computeAccessKey(CallInst *Call, 991 CallInfo &CInfo, 992 std::string &AccessKey, 993 bool &IsInt32Ret) { 994 DIType *Ty = stripQualifiers(cast<DIType>(CInfo.Metadata), false); 995 assert(!Ty->getName().empty()); 996 997 int64_t PatchImm; 998 std::string AccessStr("0"); 999 if (CInfo.AccessIndex == BTF::TYPE_EXISTENCE || 1000 CInfo.AccessIndex == BTF::TYPE_MATCH) { 1001 PatchImm = 1; 1002 } else if (CInfo.AccessIndex == BTF::TYPE_SIZE) { 1003 // typedef debuginfo type has size 0, get the eventual base type. 1004 DIType *BaseTy = stripQualifiers(Ty, true); 1005 PatchImm = BaseTy->getSizeInBits() / 8; 1006 } else { 1007 // ENUM_VALUE_EXISTENCE and ENUM_VALUE 1008 IsInt32Ret = false; 1009 1010 // The argument could be a global variable or a getelementptr with base to 1011 // a global variable depending on whether the clang option `opaque-options` 1012 // is set or not. 1013 const GlobalVariable *GV = 1014 cast<GlobalVariable>(Call->getArgOperand(1)->stripPointerCasts()); 1015 assert(GV->hasInitializer()); 1016 const ConstantDataArray *DA = cast<ConstantDataArray>(GV->getInitializer()); 1017 assert(DA->isString()); 1018 StringRef ValueStr = DA->getAsString(); 1019 1020 // ValueStr format: <EnumeratorStr>:<Value> 1021 size_t Separator = ValueStr.find_first_of(':'); 1022 StringRef EnumeratorStr = ValueStr.substr(0, Separator); 1023 1024 // Find enumerator index in the debuginfo 1025 DIType *BaseTy = stripQualifiers(Ty, true); 1026 const auto *CTy = cast<DICompositeType>(BaseTy); 1027 assert(CTy->getTag() == dwarf::DW_TAG_enumeration_type); 1028 int EnumIndex = 0; 1029 for (const auto Element : CTy->getElements()) { 1030 const auto *Enum = cast<DIEnumerator>(Element); 1031 if (Enum->getName() == EnumeratorStr) { 1032 AccessStr = std::to_string(EnumIndex); 1033 break; 1034 } 1035 EnumIndex++; 1036 } 1037 1038 if (CInfo.AccessIndex == BTF::ENUM_VALUE) { 1039 StringRef EValueStr = ValueStr.substr(Separator + 1); 1040 PatchImm = std::stoll(std::string(EValueStr)); 1041 } else { 1042 PatchImm = 1; 1043 } 1044 } 1045 1046 AccessKey = "llvm." + Ty->getName().str() + ":" + 1047 std::to_string(CInfo.AccessIndex) + std::string(":") + 1048 std::to_string(PatchImm) + std::string("$") + AccessStr; 1049 1050 return Ty; 1051 } 1052 1053 /// Call/Kind is the base preserve_*_access_index() call. Attempts to do 1054 /// transformation to a chain of relocable GEPs. 1055 bool BPFAbstractMemberAccess::transformGEPChain(CallInst *Call, 1056 CallInfo &CInfo) { 1057 std::string AccessKey; 1058 MDNode *TypeMeta; 1059 Value *Base = nullptr; 1060 bool IsInt32Ret; 1061 1062 IsInt32Ret = CInfo.Kind == BPFPreserveFieldInfoAI; 1063 if (CInfo.Kind == BPFPreserveFieldInfoAI && CInfo.Metadata) { 1064 TypeMeta = computeAccessKey(Call, CInfo, AccessKey, IsInt32Ret); 1065 } else { 1066 Base = computeBaseAndAccessKey(Call, CInfo, AccessKey, TypeMeta); 1067 if (!Base) 1068 return false; 1069 } 1070 1071 BasicBlock *BB = Call->getParent(); 1072 GlobalVariable *GV; 1073 1074 if (GEPGlobals.find(AccessKey) == GEPGlobals.end()) { 1075 IntegerType *VarType; 1076 if (IsInt32Ret) 1077 VarType = Type::getInt32Ty(BB->getContext()); // 32bit return value 1078 else 1079 VarType = Type::getInt64Ty(BB->getContext()); // 64bit ptr or enum value 1080 1081 GV = new GlobalVariable(*M, VarType, false, GlobalVariable::ExternalLinkage, 1082 nullptr, AccessKey); 1083 GV->addAttribute(BPFCoreSharedInfo::AmaAttr); 1084 GV->setMetadata(LLVMContext::MD_preserve_access_index, TypeMeta); 1085 GEPGlobals[AccessKey] = GV; 1086 } else { 1087 GV = GEPGlobals[AccessKey]; 1088 } 1089 1090 if (CInfo.Kind == BPFPreserveFieldInfoAI) { 1091 // Load the global variable which represents the returned field info. 1092 LoadInst *LDInst; 1093 if (IsInt32Ret) 1094 LDInst = new LoadInst(Type::getInt32Ty(BB->getContext()), GV, "", Call); 1095 else 1096 LDInst = new LoadInst(Type::getInt64Ty(BB->getContext()), GV, "", Call); 1097 1098 Instruction *PassThroughInst = 1099 BPFCoreSharedInfo::insertPassThrough(M, BB, LDInst, Call); 1100 Call->replaceAllUsesWith(PassThroughInst); 1101 Call->eraseFromParent(); 1102 return true; 1103 } 1104 1105 // For any original GEP Call and Base %2 like 1106 // %4 = bitcast %struct.net_device** %dev1 to i64* 1107 // it is transformed to: 1108 // %6 = load llvm.sk_buff:0:50$0:0:0:2:0 1109 // %7 = bitcast %struct.sk_buff* %2 to i8* 1110 // %8 = getelementptr i8, i8* %7, %6 1111 // %9 = bitcast i8* %8 to i64* 1112 // using %9 instead of %4 1113 // The original Call inst is removed. 1114 1115 // Load the global variable. 1116 auto *LDInst = new LoadInst(Type::getInt64Ty(BB->getContext()), GV, "", Call); 1117 1118 // Generate a BitCast 1119 auto *BCInst = 1120 new BitCastInst(Base, PointerType::getUnqual(BB->getContext())); 1121 BCInst->insertBefore(Call); 1122 1123 // Generate a GetElementPtr 1124 auto *GEP = GetElementPtrInst::Create(Type::getInt8Ty(BB->getContext()), 1125 BCInst, LDInst); 1126 GEP->insertBefore(Call); 1127 1128 // Generate a BitCast 1129 auto *BCInst2 = new BitCastInst(GEP, Call->getType()); 1130 BCInst2->insertBefore(Call); 1131 1132 // For the following code, 1133 // Block0: 1134 // ... 1135 // if (...) goto Block1 else ... 1136 // Block1: 1137 // %6 = load llvm.sk_buff:0:50$0:0:0:2:0 1138 // %7 = bitcast %struct.sk_buff* %2 to i8* 1139 // %8 = getelementptr i8, i8* %7, %6 1140 // ... 1141 // goto CommonExit 1142 // Block2: 1143 // ... 1144 // if (...) goto Block3 else ... 1145 // Block3: 1146 // %6 = load llvm.bpf_map:0:40$0:0:0:2:0 1147 // %7 = bitcast %struct.sk_buff* %2 to i8* 1148 // %8 = getelementptr i8, i8* %7, %6 1149 // ... 1150 // goto CommonExit 1151 // CommonExit 1152 // SimplifyCFG may generate: 1153 // Block0: 1154 // ... 1155 // if (...) goto Block_Common else ... 1156 // Block2: 1157 // ... 1158 // if (...) goto Block_Common else ... 1159 // Block_Common: 1160 // PHI = [llvm.sk_buff:0:50$0:0:0:2:0, llvm.bpf_map:0:40$0:0:0:2:0] 1161 // %6 = load PHI 1162 // %7 = bitcast %struct.sk_buff* %2 to i8* 1163 // %8 = getelementptr i8, i8* %7, %6 1164 // ... 1165 // goto CommonExit 1166 // For the above code, we cannot perform proper relocation since 1167 // "load PHI" has two possible relocations. 1168 // 1169 // To prevent above tail merging, we use __builtin_bpf_passthrough() 1170 // where one of its parameters is a seq_num. Since two 1171 // __builtin_bpf_passthrough() funcs will always have different seq_num, 1172 // tail merging cannot happen. The __builtin_bpf_passthrough() will be 1173 // removed in the beginning of Target IR passes. 1174 // 1175 // This approach is also used in other places when global var 1176 // representing a relocation is used. 1177 Instruction *PassThroughInst = 1178 BPFCoreSharedInfo::insertPassThrough(M, BB, BCInst2, Call); 1179 Call->replaceAllUsesWith(PassThroughInst); 1180 Call->eraseFromParent(); 1181 1182 return true; 1183 } 1184 1185 bool BPFAbstractMemberAccess::doTransformation(Function &F) { 1186 bool Transformed = false; 1187 1188 // Collect PreserveDIAccessIndex Intrinsic call chains. 1189 // The call chains will be used to generate the access 1190 // patterns similar to GEP. 1191 collectAICallChains(F); 1192 1193 for (auto &C : BaseAICalls) 1194 Transformed = transformGEPChain(C.first, C.second) || Transformed; 1195 1196 return removePreserveAccessIndexIntrinsic(F) || Transformed; 1197 } 1198 1199 PreservedAnalyses 1200 BPFAbstractMemberAccessPass::run(Function &F, FunctionAnalysisManager &AM) { 1201 return BPFAbstractMemberAccess(TM).run(F) ? PreservedAnalyses::none() 1202 : PreservedAnalyses::all(); 1203 } 1204