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/IR/DebugInfoMetadata.h" 82 #include "llvm/IR/GlobalVariable.h" 83 #include "llvm/IR/Instruction.h" 84 #include "llvm/IR/Instructions.h" 85 #include "llvm/IR/IntrinsicsBPF.h" 86 #include "llvm/IR/Module.h" 87 #include "llvm/IR/PassManager.h" 88 #include "llvm/IR/Type.h" 89 #include "llvm/IR/User.h" 90 #include "llvm/IR/Value.h" 91 #include "llvm/Pass.h" 92 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 93 #include <stack> 94 95 #define DEBUG_TYPE "bpf-abstract-member-access" 96 97 namespace llvm { 98 constexpr StringRef BPFCoreSharedInfo::AmaAttr; 99 uint32_t BPFCoreSharedInfo::SeqNum; 100 101 Instruction *BPFCoreSharedInfo::insertPassThrough(Module *M, BasicBlock *BB, 102 Instruction *Input, 103 Instruction *Before) { 104 Function *Fn = Intrinsic::getDeclaration( 105 M, Intrinsic::bpf_passthrough, {Input->getType(), Input->getType()}); 106 Constant *SeqNumVal = ConstantInt::get(Type::getInt32Ty(BB->getContext()), 107 BPFCoreSharedInfo::SeqNum++); 108 109 auto *NewInst = CallInst::Create(Fn, {SeqNumVal, Input}); 110 NewInst->insertBefore(Before); 111 return NewInst; 112 } 113 } // namespace llvm 114 115 using namespace llvm; 116 117 namespace { 118 class BPFAbstractMemberAccess final { 119 public: 120 BPFAbstractMemberAccess(BPFTargetMachine *TM) : TM(TM) {} 121 122 bool run(Function &F); 123 124 struct CallInfo { 125 uint32_t Kind; 126 uint32_t AccessIndex; 127 MaybeAlign RecordAlignment; 128 MDNode *Metadata; 129 WeakTrackingVH Base; 130 }; 131 typedef std::stack<std::pair<CallInst *, CallInfo>> CallInfoStack; 132 133 private: 134 enum : uint32_t { 135 BPFPreserveArrayAI = 1, 136 BPFPreserveUnionAI = 2, 137 BPFPreserveStructAI = 3, 138 BPFPreserveFieldInfoAI = 4, 139 }; 140 141 TargetMachine *TM; 142 const DataLayout *DL = nullptr; 143 Module *M = nullptr; 144 145 static std::map<std::string, GlobalVariable *> GEPGlobals; 146 // A map to link preserve_*_access_index intrinsic calls. 147 std::map<CallInst *, std::pair<CallInst *, CallInfo>> AIChain; 148 // A map to hold all the base preserve_*_access_index intrinsic calls. 149 // The base call is not an input of any other preserve_* 150 // intrinsics. 151 std::map<CallInst *, CallInfo> BaseAICalls; 152 // A map to hold <AnonRecord, TypeDef> relationships 153 std::map<DICompositeType *, DIDerivedType *> AnonRecords; 154 155 void CheckAnonRecordType(DIDerivedType *ParentTy, DIType *Ty); 156 void CheckCompositeType(DIDerivedType *ParentTy, DICompositeType *CTy); 157 void CheckDerivedType(DIDerivedType *ParentTy, DIDerivedType *DTy); 158 void ResetMetadata(struct CallInfo &CInfo); 159 160 bool doTransformation(Function &F); 161 162 void traceAICall(CallInst *Call, CallInfo &ParentInfo); 163 void traceBitCast(BitCastInst *BitCast, CallInst *Parent, 164 CallInfo &ParentInfo); 165 void traceGEP(GetElementPtrInst *GEP, CallInst *Parent, 166 CallInfo &ParentInfo); 167 void collectAICallChains(Function &F); 168 169 bool IsPreserveDIAccessIndexCall(const CallInst *Call, CallInfo &Cinfo); 170 bool IsValidAIChain(const MDNode *ParentMeta, uint32_t ParentAI, 171 const MDNode *ChildMeta); 172 bool removePreserveAccessIndexIntrinsic(Function &F); 173 void replaceWithGEP(std::vector<CallInst *> &CallList, 174 uint32_t NumOfZerosIndex, uint32_t DIIndex); 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 uint64_t getConstant(const Value *IndexValue); 187 bool transformGEPChain(CallInst *Call, CallInfo &CInfo); 188 }; 189 190 std::map<std::string, GlobalVariable *> BPFAbstractMemberAccess::GEPGlobals; 191 192 class BPFAbstractMemberAccessLegacyPass final : public FunctionPass { 193 BPFTargetMachine *TM; 194 195 bool runOnFunction(Function &F) override { 196 return BPFAbstractMemberAccess(TM).run(F); 197 } 198 199 public: 200 static char ID; 201 202 // Add optional BPFTargetMachine parameter so that BPF backend can add the 203 // phase with target machine to find out the endianness. The default 204 // constructor (without parameters) is used by the pass manager for managing 205 // purposes. 206 BPFAbstractMemberAccessLegacyPass(BPFTargetMachine *TM = nullptr) 207 : FunctionPass(ID), TM(TM) {} 208 }; 209 210 } // End anonymous namespace 211 212 char BPFAbstractMemberAccessLegacyPass::ID = 0; 213 INITIALIZE_PASS(BPFAbstractMemberAccessLegacyPass, DEBUG_TYPE, 214 "BPF Abstract Member Access", false, false) 215 216 FunctionPass *llvm::createBPFAbstractMemberAccess(BPFTargetMachine *TM) { 217 return new BPFAbstractMemberAccessLegacyPass(TM); 218 } 219 220 bool BPFAbstractMemberAccess::run(Function &F) { 221 LLVM_DEBUG(dbgs() << "********** Abstract Member Accesses **********\n"); 222 223 M = F.getParent(); 224 if (!M) 225 return false; 226 227 // Bail out if no debug info. 228 if (M->debug_compile_units().empty()) 229 return false; 230 231 // For each argument/return/local_variable type, trace the type 232 // pattern like '[derived_type]* [composite_type]' to check 233 // and remember (anon record -> typedef) relations where the 234 // anon record is defined as 235 // typedef [const/volatile/restrict]* [anon record] 236 DISubprogram *SP = F.getSubprogram(); 237 if (SP && SP->isDefinition()) { 238 for (DIType *Ty: SP->getType()->getTypeArray()) 239 CheckAnonRecordType(nullptr, Ty); 240 for (const DINode *DN : SP->getRetainedNodes()) { 241 if (const auto *DV = dyn_cast<DILocalVariable>(DN)) 242 CheckAnonRecordType(nullptr, DV->getType()); 243 } 244 } 245 246 DL = &M->getDataLayout(); 247 return doTransformation(F); 248 } 249 250 void BPFAbstractMemberAccess::ResetMetadata(struct CallInfo &CInfo) { 251 if (auto Ty = dyn_cast<DICompositeType>(CInfo.Metadata)) { 252 if (AnonRecords.find(Ty) != AnonRecords.end()) { 253 if (AnonRecords[Ty] != nullptr) 254 CInfo.Metadata = AnonRecords[Ty]; 255 } 256 } 257 } 258 259 void BPFAbstractMemberAccess::CheckCompositeType(DIDerivedType *ParentTy, 260 DICompositeType *CTy) { 261 if (!CTy->getName().empty() || !ParentTy || 262 ParentTy->getTag() != dwarf::DW_TAG_typedef) 263 return; 264 265 if (AnonRecords.find(CTy) == AnonRecords.end()) { 266 AnonRecords[CTy] = ParentTy; 267 return; 268 } 269 270 // Two or more typedef's may point to the same anon record. 271 // If this is the case, set the typedef DIType to be nullptr 272 // to indicate the duplication case. 273 DIDerivedType *CurrTy = AnonRecords[CTy]; 274 if (CurrTy == ParentTy) 275 return; 276 AnonRecords[CTy] = nullptr; 277 } 278 279 void BPFAbstractMemberAccess::CheckDerivedType(DIDerivedType *ParentTy, 280 DIDerivedType *DTy) { 281 DIType *BaseType = DTy->getBaseType(); 282 if (!BaseType) 283 return; 284 285 unsigned Tag = DTy->getTag(); 286 if (Tag == dwarf::DW_TAG_pointer_type) 287 CheckAnonRecordType(nullptr, BaseType); 288 else if (Tag == dwarf::DW_TAG_typedef) 289 CheckAnonRecordType(DTy, BaseType); 290 else 291 CheckAnonRecordType(ParentTy, BaseType); 292 } 293 294 void BPFAbstractMemberAccess::CheckAnonRecordType(DIDerivedType *ParentTy, 295 DIType *Ty) { 296 if (!Ty) 297 return; 298 299 if (auto *CTy = dyn_cast<DICompositeType>(Ty)) 300 return CheckCompositeType(ParentTy, CTy); 301 else if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) 302 return CheckDerivedType(ParentTy, DTy); 303 } 304 305 static bool SkipDIDerivedTag(unsigned Tag, bool skipTypedef) { 306 if (Tag != dwarf::DW_TAG_typedef && Tag != dwarf::DW_TAG_const_type && 307 Tag != dwarf::DW_TAG_volatile_type && 308 Tag != dwarf::DW_TAG_restrict_type && 309 Tag != dwarf::DW_TAG_member) 310 return false; 311 if (Tag == dwarf::DW_TAG_typedef && !skipTypedef) 312 return false; 313 return true; 314 } 315 316 static DIType * stripQualifiers(DIType *Ty, bool skipTypedef = true) { 317 while (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 318 if (!SkipDIDerivedTag(DTy->getTag(), skipTypedef)) 319 break; 320 Ty = DTy->getBaseType(); 321 } 322 return Ty; 323 } 324 325 static const DIType * stripQualifiers(const DIType *Ty) { 326 while (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 327 if (!SkipDIDerivedTag(DTy->getTag(), true)) 328 break; 329 Ty = DTy->getBaseType(); 330 } 331 return Ty; 332 } 333 334 static uint32_t calcArraySize(const DICompositeType *CTy, uint32_t StartDim) { 335 DINodeArray Elements = CTy->getElements(); 336 uint32_t DimSize = 1; 337 for (uint32_t I = StartDim; I < Elements.size(); ++I) { 338 if (auto *Element = dyn_cast_or_null<DINode>(Elements[I])) 339 if (Element->getTag() == dwarf::DW_TAG_subrange_type) { 340 const DISubrange *SR = cast<DISubrange>(Element); 341 auto *CI = SR->getCount().dyn_cast<ConstantInt *>(); 342 DimSize *= CI->getSExtValue(); 343 } 344 } 345 346 return DimSize; 347 } 348 349 static Type *getBaseElementType(const CallInst *Call) { 350 // Element type is stored in an elementtype() attribute on the first param. 351 return Call->getParamElementType(0); 352 } 353 354 /// Check whether a call is a preserve_*_access_index intrinsic call or not. 355 bool BPFAbstractMemberAccess::IsPreserveDIAccessIndexCall(const CallInst *Call, 356 CallInfo &CInfo) { 357 if (!Call) 358 return false; 359 360 const auto *GV = dyn_cast<GlobalValue>(Call->getCalledOperand()); 361 if (!GV) 362 return false; 363 if (GV->getName().startswith("llvm.preserve.array.access.index")) { 364 CInfo.Kind = BPFPreserveArrayAI; 365 CInfo.Metadata = Call->getMetadata(LLVMContext::MD_preserve_access_index); 366 if (!CInfo.Metadata) 367 report_fatal_error("Missing metadata for llvm.preserve.array.access.index intrinsic"); 368 CInfo.AccessIndex = getConstant(Call->getArgOperand(2)); 369 CInfo.Base = Call->getArgOperand(0); 370 CInfo.RecordAlignment = DL->getABITypeAlign(getBaseElementType(Call)); 371 return true; 372 } 373 if (GV->getName().startswith("llvm.preserve.union.access.index")) { 374 CInfo.Kind = BPFPreserveUnionAI; 375 CInfo.Metadata = Call->getMetadata(LLVMContext::MD_preserve_access_index); 376 if (!CInfo.Metadata) 377 report_fatal_error("Missing metadata for llvm.preserve.union.access.index intrinsic"); 378 ResetMetadata(CInfo); 379 CInfo.AccessIndex = getConstant(Call->getArgOperand(1)); 380 CInfo.Base = Call->getArgOperand(0); 381 return true; 382 } 383 if (GV->getName().startswith("llvm.preserve.struct.access.index")) { 384 CInfo.Kind = BPFPreserveStructAI; 385 CInfo.Metadata = Call->getMetadata(LLVMContext::MD_preserve_access_index); 386 if (!CInfo.Metadata) 387 report_fatal_error("Missing metadata for llvm.preserve.struct.access.index intrinsic"); 388 ResetMetadata(CInfo); 389 CInfo.AccessIndex = getConstant(Call->getArgOperand(2)); 390 CInfo.Base = Call->getArgOperand(0); 391 CInfo.RecordAlignment = DL->getABITypeAlign(getBaseElementType(Call)); 392 return true; 393 } 394 if (GV->getName().startswith("llvm.bpf.preserve.field.info")) { 395 CInfo.Kind = BPFPreserveFieldInfoAI; 396 CInfo.Metadata = nullptr; 397 // Check validity of info_kind as clang did not check this. 398 uint64_t InfoKind = getConstant(Call->getArgOperand(1)); 399 if (InfoKind >= BPFCoreSharedInfo::MAX_FIELD_RELOC_KIND) 400 report_fatal_error("Incorrect info_kind for llvm.bpf.preserve.field.info intrinsic"); 401 CInfo.AccessIndex = InfoKind; 402 return true; 403 } 404 if (GV->getName().startswith("llvm.bpf.preserve.type.info")) { 405 CInfo.Kind = BPFPreserveFieldInfoAI; 406 CInfo.Metadata = Call->getMetadata(LLVMContext::MD_preserve_access_index); 407 if (!CInfo.Metadata) 408 report_fatal_error("Missing metadata for llvm.preserve.type.info intrinsic"); 409 uint64_t Flag = getConstant(Call->getArgOperand(1)); 410 if (Flag >= BPFCoreSharedInfo::MAX_PRESERVE_TYPE_INFO_FLAG) 411 report_fatal_error("Incorrect flag for llvm.bpf.preserve.type.info intrinsic"); 412 if (Flag == BPFCoreSharedInfo::PRESERVE_TYPE_INFO_EXISTENCE) 413 CInfo.AccessIndex = BPFCoreSharedInfo::TYPE_EXISTENCE; 414 else if (Flag == BPFCoreSharedInfo::PRESERVE_TYPE_INFO_MATCH) 415 CInfo.AccessIndex = BPFCoreSharedInfo::TYPE_MATCH; 416 else 417 CInfo.AccessIndex = BPFCoreSharedInfo::TYPE_SIZE; 418 return true; 419 } 420 if (GV->getName().startswith("llvm.bpf.preserve.enum.value")) { 421 CInfo.Kind = BPFPreserveFieldInfoAI; 422 CInfo.Metadata = Call->getMetadata(LLVMContext::MD_preserve_access_index); 423 if (!CInfo.Metadata) 424 report_fatal_error("Missing metadata for llvm.preserve.enum.value intrinsic"); 425 uint64_t Flag = getConstant(Call->getArgOperand(2)); 426 if (Flag >= BPFCoreSharedInfo::MAX_PRESERVE_ENUM_VALUE_FLAG) 427 report_fatal_error("Incorrect flag for llvm.bpf.preserve.enum.value intrinsic"); 428 if (Flag == BPFCoreSharedInfo::PRESERVE_ENUM_VALUE_EXISTENCE) 429 CInfo.AccessIndex = BPFCoreSharedInfo::ENUM_VALUE_EXISTENCE; 430 else 431 CInfo.AccessIndex = BPFCoreSharedInfo::ENUM_VALUE; 432 return true; 433 } 434 435 return false; 436 } 437 438 void BPFAbstractMemberAccess::replaceWithGEP(std::vector<CallInst *> &CallList, 439 uint32_t DimensionIndex, 440 uint32_t GEPIndex) { 441 for (auto *Call : CallList) { 442 uint32_t Dimension = 1; 443 if (DimensionIndex > 0) 444 Dimension = getConstant(Call->getArgOperand(DimensionIndex)); 445 446 Constant *Zero = 447 ConstantInt::get(Type::getInt32Ty(Call->getParent()->getContext()), 0); 448 SmallVector<Value *, 4> IdxList; 449 for (unsigned I = 0; I < Dimension; ++I) 450 IdxList.push_back(Zero); 451 IdxList.push_back(Call->getArgOperand(GEPIndex)); 452 453 auto *GEP = GetElementPtrInst::CreateInBounds( 454 getBaseElementType(Call), Call->getArgOperand(0), IdxList, "", Call); 455 Call->replaceAllUsesWith(GEP); 456 Call->eraseFromParent(); 457 } 458 } 459 460 bool BPFAbstractMemberAccess::removePreserveAccessIndexIntrinsic(Function &F) { 461 std::vector<CallInst *> PreserveArrayIndexCalls; 462 std::vector<CallInst *> PreserveUnionIndexCalls; 463 std::vector<CallInst *> PreserveStructIndexCalls; 464 bool Found = false; 465 466 for (auto &BB : F) 467 for (auto &I : BB) { 468 auto *Call = dyn_cast<CallInst>(&I); 469 CallInfo CInfo; 470 if (!IsPreserveDIAccessIndexCall(Call, CInfo)) 471 continue; 472 473 Found = true; 474 if (CInfo.Kind == BPFPreserveArrayAI) 475 PreserveArrayIndexCalls.push_back(Call); 476 else if (CInfo.Kind == BPFPreserveUnionAI) 477 PreserveUnionIndexCalls.push_back(Call); 478 else 479 PreserveStructIndexCalls.push_back(Call); 480 } 481 482 // do the following transformation: 483 // . addr = preserve_array_access_index(base, dimension, index) 484 // is transformed to 485 // addr = GEP(base, dimenion's zero's, index) 486 // . addr = preserve_union_access_index(base, di_index) 487 // is transformed to 488 // addr = base, i.e., all usages of "addr" are replaced by "base". 489 // . addr = preserve_struct_access_index(base, gep_index, di_index) 490 // is transformed to 491 // addr = GEP(base, 0, gep_index) 492 replaceWithGEP(PreserveArrayIndexCalls, 1, 2); 493 replaceWithGEP(PreserveStructIndexCalls, 0, 1); 494 for (auto *Call : PreserveUnionIndexCalls) { 495 Call->replaceAllUsesWith(Call->getArgOperand(0)); 496 Call->eraseFromParent(); 497 } 498 499 return Found; 500 } 501 502 /// Check whether the access index chain is valid. We check 503 /// here because there may be type casts between two 504 /// access indexes. We want to ensure memory access still valid. 505 bool BPFAbstractMemberAccess::IsValidAIChain(const MDNode *ParentType, 506 uint32_t ParentAI, 507 const MDNode *ChildType) { 508 if (!ChildType) 509 return true; // preserve_field_info, no type comparison needed. 510 511 const DIType *PType = stripQualifiers(cast<DIType>(ParentType)); 512 const DIType *CType = stripQualifiers(cast<DIType>(ChildType)); 513 514 // Child is a derived/pointer type, which is due to type casting. 515 // Pointer type cannot be in the middle of chain. 516 if (isa<DIDerivedType>(CType)) 517 return false; 518 519 // Parent is a pointer type. 520 if (const auto *PtrTy = dyn_cast<DIDerivedType>(PType)) { 521 if (PtrTy->getTag() != dwarf::DW_TAG_pointer_type) 522 return false; 523 return stripQualifiers(PtrTy->getBaseType()) == CType; 524 } 525 526 // Otherwise, struct/union/array types 527 const auto *PTy = dyn_cast<DICompositeType>(PType); 528 const auto *CTy = dyn_cast<DICompositeType>(CType); 529 assert(PTy && CTy && "ParentType or ChildType is null or not composite"); 530 531 uint32_t PTyTag = PTy->getTag(); 532 assert(PTyTag == dwarf::DW_TAG_array_type || 533 PTyTag == dwarf::DW_TAG_structure_type || 534 PTyTag == dwarf::DW_TAG_union_type); 535 536 uint32_t CTyTag = CTy->getTag(); 537 assert(CTyTag == dwarf::DW_TAG_array_type || 538 CTyTag == dwarf::DW_TAG_structure_type || 539 CTyTag == dwarf::DW_TAG_union_type); 540 541 // Multi dimensional arrays, base element should be the same 542 if (PTyTag == dwarf::DW_TAG_array_type && PTyTag == CTyTag) 543 return PTy->getBaseType() == CTy->getBaseType(); 544 545 DIType *Ty; 546 if (PTyTag == dwarf::DW_TAG_array_type) 547 Ty = PTy->getBaseType(); 548 else 549 Ty = dyn_cast<DIType>(PTy->getElements()[ParentAI]); 550 551 return dyn_cast<DICompositeType>(stripQualifiers(Ty)) == CTy; 552 } 553 554 void BPFAbstractMemberAccess::traceAICall(CallInst *Call, 555 CallInfo &ParentInfo) { 556 for (User *U : Call->users()) { 557 Instruction *Inst = dyn_cast<Instruction>(U); 558 if (!Inst) 559 continue; 560 561 if (auto *BI = dyn_cast<BitCastInst>(Inst)) { 562 traceBitCast(BI, Call, ParentInfo); 563 } else if (auto *CI = dyn_cast<CallInst>(Inst)) { 564 CallInfo ChildInfo; 565 566 if (IsPreserveDIAccessIndexCall(CI, ChildInfo) && 567 IsValidAIChain(ParentInfo.Metadata, ParentInfo.AccessIndex, 568 ChildInfo.Metadata)) { 569 AIChain[CI] = std::make_pair(Call, ParentInfo); 570 traceAICall(CI, ChildInfo); 571 } else { 572 BaseAICalls[Call] = ParentInfo; 573 } 574 } else if (auto *GI = dyn_cast<GetElementPtrInst>(Inst)) { 575 if (GI->hasAllZeroIndices()) 576 traceGEP(GI, Call, ParentInfo); 577 else 578 BaseAICalls[Call] = ParentInfo; 579 } else { 580 BaseAICalls[Call] = ParentInfo; 581 } 582 } 583 } 584 585 void BPFAbstractMemberAccess::traceBitCast(BitCastInst *BitCast, 586 CallInst *Parent, 587 CallInfo &ParentInfo) { 588 for (User *U : BitCast->users()) { 589 Instruction *Inst = dyn_cast<Instruction>(U); 590 if (!Inst) 591 continue; 592 593 if (auto *BI = dyn_cast<BitCastInst>(Inst)) { 594 traceBitCast(BI, Parent, ParentInfo); 595 } else if (auto *CI = dyn_cast<CallInst>(Inst)) { 596 CallInfo ChildInfo; 597 if (IsPreserveDIAccessIndexCall(CI, ChildInfo) && 598 IsValidAIChain(ParentInfo.Metadata, ParentInfo.AccessIndex, 599 ChildInfo.Metadata)) { 600 AIChain[CI] = std::make_pair(Parent, ParentInfo); 601 traceAICall(CI, ChildInfo); 602 } else { 603 BaseAICalls[Parent] = ParentInfo; 604 } 605 } else if (auto *GI = dyn_cast<GetElementPtrInst>(Inst)) { 606 if (GI->hasAllZeroIndices()) 607 traceGEP(GI, Parent, ParentInfo); 608 else 609 BaseAICalls[Parent] = ParentInfo; 610 } else { 611 BaseAICalls[Parent] = ParentInfo; 612 } 613 } 614 } 615 616 void BPFAbstractMemberAccess::traceGEP(GetElementPtrInst *GEP, CallInst *Parent, 617 CallInfo &ParentInfo) { 618 for (User *U : GEP->users()) { 619 Instruction *Inst = dyn_cast<Instruction>(U); 620 if (!Inst) 621 continue; 622 623 if (auto *BI = dyn_cast<BitCastInst>(Inst)) { 624 traceBitCast(BI, Parent, ParentInfo); 625 } else if (auto *CI = dyn_cast<CallInst>(Inst)) { 626 CallInfo ChildInfo; 627 if (IsPreserveDIAccessIndexCall(CI, ChildInfo) && 628 IsValidAIChain(ParentInfo.Metadata, ParentInfo.AccessIndex, 629 ChildInfo.Metadata)) { 630 AIChain[CI] = std::make_pair(Parent, ParentInfo); 631 traceAICall(CI, ChildInfo); 632 } else { 633 BaseAICalls[Parent] = ParentInfo; 634 } 635 } else if (auto *GI = dyn_cast<GetElementPtrInst>(Inst)) { 636 if (GI->hasAllZeroIndices()) 637 traceGEP(GI, Parent, ParentInfo); 638 else 639 BaseAICalls[Parent] = ParentInfo; 640 } else { 641 BaseAICalls[Parent] = ParentInfo; 642 } 643 } 644 } 645 646 void BPFAbstractMemberAccess::collectAICallChains(Function &F) { 647 AIChain.clear(); 648 BaseAICalls.clear(); 649 650 for (auto &BB : F) 651 for (auto &I : BB) { 652 CallInfo CInfo; 653 auto *Call = dyn_cast<CallInst>(&I); 654 if (!IsPreserveDIAccessIndexCall(Call, CInfo) || 655 AIChain.find(Call) != AIChain.end()) 656 continue; 657 658 traceAICall(Call, CInfo); 659 } 660 } 661 662 uint64_t BPFAbstractMemberAccess::getConstant(const Value *IndexValue) { 663 const ConstantInt *CV = dyn_cast<ConstantInt>(IndexValue); 664 assert(CV); 665 return CV->getValue().getZExtValue(); 666 } 667 668 /// Get the start and the end of storage offset for \p MemberTy. 669 void BPFAbstractMemberAccess::GetStorageBitRange(DIDerivedType *MemberTy, 670 Align RecordAlignment, 671 uint32_t &StartBitOffset, 672 uint32_t &EndBitOffset) { 673 uint32_t MemberBitSize = MemberTy->getSizeInBits(); 674 uint32_t MemberBitOffset = MemberTy->getOffsetInBits(); 675 676 if (RecordAlignment > 8) { 677 // If the Bits are within an aligned 8-byte, set the RecordAlignment 678 // to 8, other report the fatal error. 679 if (MemberBitOffset / 64 != (MemberBitOffset + MemberBitSize) / 64) 680 report_fatal_error("Unsupported field expression for llvm.bpf.preserve.field.info, " 681 "requiring too big alignment"); 682 RecordAlignment = Align(8); 683 } 684 685 uint32_t AlignBits = RecordAlignment.value() * 8; 686 if (MemberBitSize > AlignBits) 687 report_fatal_error("Unsupported field expression for llvm.bpf.preserve.field.info, " 688 "bitfield size greater than record alignment"); 689 690 StartBitOffset = MemberBitOffset & ~(AlignBits - 1); 691 if ((StartBitOffset + AlignBits) < (MemberBitOffset + MemberBitSize)) 692 report_fatal_error("Unsupported field expression for llvm.bpf.preserve.field.info, " 693 "cross alignment boundary"); 694 EndBitOffset = StartBitOffset + AlignBits; 695 } 696 697 uint32_t BPFAbstractMemberAccess::GetFieldInfo(uint32_t InfoKind, 698 DICompositeType *CTy, 699 uint32_t AccessIndex, 700 uint32_t PatchImm, 701 MaybeAlign RecordAlignment) { 702 if (InfoKind == BPFCoreSharedInfo::FIELD_EXISTENCE) 703 return 1; 704 705 uint32_t Tag = CTy->getTag(); 706 if (InfoKind == BPFCoreSharedInfo::FIELD_BYTE_OFFSET) { 707 if (Tag == dwarf::DW_TAG_array_type) { 708 auto *EltTy = stripQualifiers(CTy->getBaseType()); 709 PatchImm += AccessIndex * calcArraySize(CTy, 1) * 710 (EltTy->getSizeInBits() >> 3); 711 } else if (Tag == dwarf::DW_TAG_structure_type) { 712 auto *MemberTy = cast<DIDerivedType>(CTy->getElements()[AccessIndex]); 713 if (!MemberTy->isBitField()) { 714 PatchImm += MemberTy->getOffsetInBits() >> 3; 715 } else { 716 unsigned SBitOffset, NextSBitOffset; 717 GetStorageBitRange(MemberTy, *RecordAlignment, SBitOffset, 718 NextSBitOffset); 719 PatchImm += SBitOffset >> 3; 720 } 721 } 722 return PatchImm; 723 } 724 725 if (InfoKind == BPFCoreSharedInfo::FIELD_BYTE_SIZE) { 726 if (Tag == dwarf::DW_TAG_array_type) { 727 auto *EltTy = stripQualifiers(CTy->getBaseType()); 728 return calcArraySize(CTy, 1) * (EltTy->getSizeInBits() >> 3); 729 } else { 730 auto *MemberTy = cast<DIDerivedType>(CTy->getElements()[AccessIndex]); 731 uint32_t SizeInBits = MemberTy->getSizeInBits(); 732 if (!MemberTy->isBitField()) 733 return SizeInBits >> 3; 734 735 unsigned SBitOffset, NextSBitOffset; 736 GetStorageBitRange(MemberTy, *RecordAlignment, SBitOffset, 737 NextSBitOffset); 738 SizeInBits = NextSBitOffset - SBitOffset; 739 if (SizeInBits & (SizeInBits - 1)) 740 report_fatal_error("Unsupported field expression for llvm.bpf.preserve.field.info"); 741 return SizeInBits >> 3; 742 } 743 } 744 745 if (InfoKind == BPFCoreSharedInfo::FIELD_SIGNEDNESS) { 746 const DIType *BaseTy; 747 if (Tag == dwarf::DW_TAG_array_type) { 748 // Signedness only checked when final array elements are accessed. 749 if (CTy->getElements().size() != 1) 750 report_fatal_error("Invalid array expression for llvm.bpf.preserve.field.info"); 751 BaseTy = stripQualifiers(CTy->getBaseType()); 752 } else { 753 auto *MemberTy = cast<DIDerivedType>(CTy->getElements()[AccessIndex]); 754 BaseTy = stripQualifiers(MemberTy->getBaseType()); 755 } 756 757 // Only basic types and enum types have signedness. 758 const auto *BTy = dyn_cast<DIBasicType>(BaseTy); 759 while (!BTy) { 760 const auto *CompTy = dyn_cast<DICompositeType>(BaseTy); 761 // Report an error if the field expression does not have signedness. 762 if (!CompTy || CompTy->getTag() != dwarf::DW_TAG_enumeration_type) 763 report_fatal_error("Invalid field expression for llvm.bpf.preserve.field.info"); 764 BaseTy = stripQualifiers(CompTy->getBaseType()); 765 BTy = dyn_cast<DIBasicType>(BaseTy); 766 } 767 uint32_t Encoding = BTy->getEncoding(); 768 return (Encoding == dwarf::DW_ATE_signed || Encoding == dwarf::DW_ATE_signed_char); 769 } 770 771 if (InfoKind == BPFCoreSharedInfo::FIELD_LSHIFT_U64) { 772 // The value is loaded into a value with FIELD_BYTE_SIZE size, 773 // and then zero or sign extended to U64. 774 // FIELD_LSHIFT_U64 and FIELD_RSHIFT_U64 are operations 775 // to extract the original value. 776 const Triple &Triple = TM->getTargetTriple(); 777 DIDerivedType *MemberTy = nullptr; 778 bool IsBitField = false; 779 uint32_t SizeInBits; 780 781 if (Tag == dwarf::DW_TAG_array_type) { 782 auto *EltTy = stripQualifiers(CTy->getBaseType()); 783 SizeInBits = calcArraySize(CTy, 1) * EltTy->getSizeInBits(); 784 } else { 785 MemberTy = cast<DIDerivedType>(CTy->getElements()[AccessIndex]); 786 SizeInBits = MemberTy->getSizeInBits(); 787 IsBitField = MemberTy->isBitField(); 788 } 789 790 if (!IsBitField) { 791 if (SizeInBits > 64) 792 report_fatal_error("too big field size for llvm.bpf.preserve.field.info"); 793 return 64 - SizeInBits; 794 } 795 796 unsigned SBitOffset, NextSBitOffset; 797 GetStorageBitRange(MemberTy, *RecordAlignment, SBitOffset, NextSBitOffset); 798 if (NextSBitOffset - SBitOffset > 64) 799 report_fatal_error("too big field size for llvm.bpf.preserve.field.info"); 800 801 unsigned OffsetInBits = MemberTy->getOffsetInBits(); 802 if (Triple.getArch() == Triple::bpfel) 803 return SBitOffset + 64 - OffsetInBits - SizeInBits; 804 else 805 return OffsetInBits + 64 - NextSBitOffset; 806 } 807 808 if (InfoKind == BPFCoreSharedInfo::FIELD_RSHIFT_U64) { 809 DIDerivedType *MemberTy = nullptr; 810 bool IsBitField = false; 811 uint32_t SizeInBits; 812 if (Tag == dwarf::DW_TAG_array_type) { 813 auto *EltTy = stripQualifiers(CTy->getBaseType()); 814 SizeInBits = calcArraySize(CTy, 1) * EltTy->getSizeInBits(); 815 } else { 816 MemberTy = cast<DIDerivedType>(CTy->getElements()[AccessIndex]); 817 SizeInBits = MemberTy->getSizeInBits(); 818 IsBitField = MemberTy->isBitField(); 819 } 820 821 if (!IsBitField) { 822 if (SizeInBits > 64) 823 report_fatal_error("too big field size for llvm.bpf.preserve.field.info"); 824 return 64 - SizeInBits; 825 } 826 827 unsigned SBitOffset, NextSBitOffset; 828 GetStorageBitRange(MemberTy, *RecordAlignment, SBitOffset, NextSBitOffset); 829 if (NextSBitOffset - SBitOffset > 64) 830 report_fatal_error("too big field size for llvm.bpf.preserve.field.info"); 831 832 return 64 - SizeInBits; 833 } 834 835 llvm_unreachable("Unknown llvm.bpf.preserve.field.info info kind"); 836 } 837 838 bool BPFAbstractMemberAccess::HasPreserveFieldInfoCall(CallInfoStack &CallStack) { 839 // This is called in error return path, no need to maintain CallStack. 840 while (CallStack.size()) { 841 auto StackElem = CallStack.top(); 842 if (StackElem.second.Kind == BPFPreserveFieldInfoAI) 843 return true; 844 CallStack.pop(); 845 } 846 return false; 847 } 848 849 /// Compute the base of the whole preserve_* intrinsics chains, i.e., the base 850 /// pointer of the first preserve_*_access_index call, and construct the access 851 /// string, which will be the name of a global variable. 852 Value *BPFAbstractMemberAccess::computeBaseAndAccessKey(CallInst *Call, 853 CallInfo &CInfo, 854 std::string &AccessKey, 855 MDNode *&TypeMeta) { 856 Value *Base = nullptr; 857 std::string TypeName; 858 CallInfoStack CallStack; 859 860 // Put the access chain into a stack with the top as the head of the chain. 861 while (Call) { 862 CallStack.push(std::make_pair(Call, CInfo)); 863 CInfo = AIChain[Call].second; 864 Call = AIChain[Call].first; 865 } 866 867 // The access offset from the base of the head of chain is also 868 // calculated here as all debuginfo types are available. 869 870 // Get type name and calculate the first index. 871 // We only want to get type name from typedef, structure or union. 872 // If user wants a relocation like 873 // int *p; ... __builtin_preserve_access_index(&p[4]) ... 874 // or 875 // int a[10][20]; ... __builtin_preserve_access_index(&a[2][3]) ... 876 // we will skip them. 877 uint32_t FirstIndex = 0; 878 uint32_t PatchImm = 0; // AccessOffset or the requested field info 879 uint32_t InfoKind = BPFCoreSharedInfo::FIELD_BYTE_OFFSET; 880 while (CallStack.size()) { 881 auto StackElem = CallStack.top(); 882 Call = StackElem.first; 883 CInfo = StackElem.second; 884 885 if (!Base) 886 Base = CInfo.Base; 887 888 DIType *PossibleTypeDef = stripQualifiers(cast<DIType>(CInfo.Metadata), 889 false); 890 DIType *Ty = stripQualifiers(PossibleTypeDef); 891 if (CInfo.Kind == BPFPreserveUnionAI || 892 CInfo.Kind == BPFPreserveStructAI) { 893 // struct or union type. If the typedef is in the metadata, always 894 // use the typedef. 895 TypeName = std::string(PossibleTypeDef->getName()); 896 TypeMeta = PossibleTypeDef; 897 PatchImm += FirstIndex * (Ty->getSizeInBits() >> 3); 898 break; 899 } 900 901 assert(CInfo.Kind == BPFPreserveArrayAI); 902 903 // Array entries will always be consumed for accumulative initial index. 904 CallStack.pop(); 905 906 // BPFPreserveArrayAI 907 uint64_t AccessIndex = CInfo.AccessIndex; 908 909 DIType *BaseTy = nullptr; 910 bool CheckElemType = false; 911 if (const auto *CTy = dyn_cast<DICompositeType>(Ty)) { 912 // array type 913 assert(CTy->getTag() == dwarf::DW_TAG_array_type); 914 915 916 FirstIndex += AccessIndex * calcArraySize(CTy, 1); 917 BaseTy = stripQualifiers(CTy->getBaseType()); 918 CheckElemType = CTy->getElements().size() == 1; 919 } else { 920 // pointer type 921 auto *DTy = cast<DIDerivedType>(Ty); 922 assert(DTy->getTag() == dwarf::DW_TAG_pointer_type); 923 924 BaseTy = stripQualifiers(DTy->getBaseType()); 925 CTy = dyn_cast<DICompositeType>(BaseTy); 926 if (!CTy) { 927 CheckElemType = true; 928 } else if (CTy->getTag() != dwarf::DW_TAG_array_type) { 929 FirstIndex += AccessIndex; 930 CheckElemType = true; 931 } else { 932 FirstIndex += AccessIndex * calcArraySize(CTy, 0); 933 } 934 } 935 936 if (CheckElemType) { 937 auto *CTy = dyn_cast<DICompositeType>(BaseTy); 938 if (!CTy) { 939 if (HasPreserveFieldInfoCall(CallStack)) 940 report_fatal_error("Invalid field access for llvm.preserve.field.info intrinsic"); 941 return nullptr; 942 } 943 944 unsigned CTag = CTy->getTag(); 945 if (CTag == dwarf::DW_TAG_structure_type || CTag == dwarf::DW_TAG_union_type) { 946 TypeName = std::string(CTy->getName()); 947 } else { 948 if (HasPreserveFieldInfoCall(CallStack)) 949 report_fatal_error("Invalid field access for llvm.preserve.field.info intrinsic"); 950 return nullptr; 951 } 952 TypeMeta = CTy; 953 PatchImm += FirstIndex * (CTy->getSizeInBits() >> 3); 954 break; 955 } 956 } 957 assert(TypeName.size()); 958 AccessKey += std::to_string(FirstIndex); 959 960 // Traverse the rest of access chain to complete offset calculation 961 // and access key construction. 962 while (CallStack.size()) { 963 auto StackElem = CallStack.top(); 964 CInfo = StackElem.second; 965 CallStack.pop(); 966 967 if (CInfo.Kind == BPFPreserveFieldInfoAI) { 968 InfoKind = CInfo.AccessIndex; 969 if (InfoKind == BPFCoreSharedInfo::FIELD_EXISTENCE) 970 PatchImm = 1; 971 break; 972 } 973 974 // If the next Call (the top of the stack) is a BPFPreserveFieldInfoAI, 975 // the action will be extracting field info. 976 if (CallStack.size()) { 977 auto StackElem2 = CallStack.top(); 978 CallInfo CInfo2 = StackElem2.second; 979 if (CInfo2.Kind == BPFPreserveFieldInfoAI) { 980 InfoKind = CInfo2.AccessIndex; 981 assert(CallStack.size() == 1); 982 } 983 } 984 985 // Access Index 986 uint64_t AccessIndex = CInfo.AccessIndex; 987 AccessKey += ":" + std::to_string(AccessIndex); 988 989 MDNode *MDN = CInfo.Metadata; 990 // At this stage, it cannot be pointer type. 991 auto *CTy = cast<DICompositeType>(stripQualifiers(cast<DIType>(MDN))); 992 PatchImm = GetFieldInfo(InfoKind, CTy, AccessIndex, PatchImm, 993 CInfo.RecordAlignment); 994 } 995 996 // Access key is the 997 // "llvm." + type name + ":" + reloc type + ":" + patched imm + "$" + 998 // access string, 999 // uniquely identifying one relocation. 1000 // The prefix "llvm." indicates this is a temporary global, which should 1001 // not be emitted to ELF file. 1002 AccessKey = "llvm." + TypeName + ":" + std::to_string(InfoKind) + ":" + 1003 std::to_string(PatchImm) + "$" + AccessKey; 1004 1005 return Base; 1006 } 1007 1008 MDNode *BPFAbstractMemberAccess::computeAccessKey(CallInst *Call, 1009 CallInfo &CInfo, 1010 std::string &AccessKey, 1011 bool &IsInt32Ret) { 1012 DIType *Ty = stripQualifiers(cast<DIType>(CInfo.Metadata), false); 1013 assert(!Ty->getName().empty()); 1014 1015 int64_t PatchImm; 1016 std::string AccessStr("0"); 1017 if (CInfo.AccessIndex == BPFCoreSharedInfo::TYPE_EXISTENCE || 1018 CInfo.AccessIndex == BPFCoreSharedInfo::TYPE_MATCH) { 1019 PatchImm = 1; 1020 } else if (CInfo.AccessIndex == BPFCoreSharedInfo::TYPE_SIZE) { 1021 // typedef debuginfo type has size 0, get the eventual base type. 1022 DIType *BaseTy = stripQualifiers(Ty, true); 1023 PatchImm = BaseTy->getSizeInBits() / 8; 1024 } else { 1025 // ENUM_VALUE_EXISTENCE and ENUM_VALUE 1026 IsInt32Ret = false; 1027 1028 // The argument could be a global variable or a getelementptr with base to 1029 // a global variable depending on whether the clang option `opaque-options` 1030 // is set or not. 1031 const GlobalVariable *GV = 1032 cast<GlobalVariable>(Call->getArgOperand(1)->stripPointerCasts()); 1033 assert(GV->hasInitializer()); 1034 const ConstantDataArray *DA = cast<ConstantDataArray>(GV->getInitializer()); 1035 assert(DA->isString()); 1036 StringRef ValueStr = DA->getAsString(); 1037 1038 // ValueStr format: <EnumeratorStr>:<Value> 1039 size_t Separator = ValueStr.find_first_of(':'); 1040 StringRef EnumeratorStr = ValueStr.substr(0, Separator); 1041 1042 // Find enumerator index in the debuginfo 1043 DIType *BaseTy = stripQualifiers(Ty, true); 1044 const auto *CTy = cast<DICompositeType>(BaseTy); 1045 assert(CTy->getTag() == dwarf::DW_TAG_enumeration_type); 1046 int EnumIndex = 0; 1047 for (const auto Element : CTy->getElements()) { 1048 const auto *Enum = cast<DIEnumerator>(Element); 1049 if (Enum->getName() == EnumeratorStr) { 1050 AccessStr = std::to_string(EnumIndex); 1051 break; 1052 } 1053 EnumIndex++; 1054 } 1055 1056 if (CInfo.AccessIndex == BPFCoreSharedInfo::ENUM_VALUE) { 1057 StringRef EValueStr = ValueStr.substr(Separator + 1); 1058 PatchImm = std::stoll(std::string(EValueStr)); 1059 } else { 1060 PatchImm = 1; 1061 } 1062 } 1063 1064 AccessKey = "llvm." + Ty->getName().str() + ":" + 1065 std::to_string(CInfo.AccessIndex) + std::string(":") + 1066 std::to_string(PatchImm) + std::string("$") + AccessStr; 1067 1068 return Ty; 1069 } 1070 1071 /// Call/Kind is the base preserve_*_access_index() call. Attempts to do 1072 /// transformation to a chain of relocable GEPs. 1073 bool BPFAbstractMemberAccess::transformGEPChain(CallInst *Call, 1074 CallInfo &CInfo) { 1075 std::string AccessKey; 1076 MDNode *TypeMeta; 1077 Value *Base = nullptr; 1078 bool IsInt32Ret; 1079 1080 IsInt32Ret = CInfo.Kind == BPFPreserveFieldInfoAI; 1081 if (CInfo.Kind == BPFPreserveFieldInfoAI && CInfo.Metadata) { 1082 TypeMeta = computeAccessKey(Call, CInfo, AccessKey, IsInt32Ret); 1083 } else { 1084 Base = computeBaseAndAccessKey(Call, CInfo, AccessKey, TypeMeta); 1085 if (!Base) 1086 return false; 1087 } 1088 1089 BasicBlock *BB = Call->getParent(); 1090 GlobalVariable *GV; 1091 1092 if (GEPGlobals.find(AccessKey) == GEPGlobals.end()) { 1093 IntegerType *VarType; 1094 if (IsInt32Ret) 1095 VarType = Type::getInt32Ty(BB->getContext()); // 32bit return value 1096 else 1097 VarType = Type::getInt64Ty(BB->getContext()); // 64bit ptr or enum value 1098 1099 GV = new GlobalVariable(*M, VarType, false, GlobalVariable::ExternalLinkage, 1100 nullptr, AccessKey); 1101 GV->addAttribute(BPFCoreSharedInfo::AmaAttr); 1102 GV->setMetadata(LLVMContext::MD_preserve_access_index, TypeMeta); 1103 GEPGlobals[AccessKey] = GV; 1104 } else { 1105 GV = GEPGlobals[AccessKey]; 1106 } 1107 1108 if (CInfo.Kind == BPFPreserveFieldInfoAI) { 1109 // Load the global variable which represents the returned field info. 1110 LoadInst *LDInst; 1111 if (IsInt32Ret) 1112 LDInst = new LoadInst(Type::getInt32Ty(BB->getContext()), GV, "", Call); 1113 else 1114 LDInst = new LoadInst(Type::getInt64Ty(BB->getContext()), GV, "", Call); 1115 1116 Instruction *PassThroughInst = 1117 BPFCoreSharedInfo::insertPassThrough(M, BB, LDInst, Call); 1118 Call->replaceAllUsesWith(PassThroughInst); 1119 Call->eraseFromParent(); 1120 return true; 1121 } 1122 1123 // For any original GEP Call and Base %2 like 1124 // %4 = bitcast %struct.net_device** %dev1 to i64* 1125 // it is transformed to: 1126 // %6 = load llvm.sk_buff:0:50$0:0:0:2:0 1127 // %7 = bitcast %struct.sk_buff* %2 to i8* 1128 // %8 = getelementptr i8, i8* %7, %6 1129 // %9 = bitcast i8* %8 to i64* 1130 // using %9 instead of %4 1131 // The original Call inst is removed. 1132 1133 // Load the global variable. 1134 auto *LDInst = new LoadInst(Type::getInt64Ty(BB->getContext()), GV, "", Call); 1135 1136 // Generate a BitCast 1137 auto *BCInst = new BitCastInst(Base, Type::getInt8PtrTy(BB->getContext())); 1138 BCInst->insertBefore(Call); 1139 1140 // Generate a GetElementPtr 1141 auto *GEP = GetElementPtrInst::Create(Type::getInt8Ty(BB->getContext()), 1142 BCInst, LDInst); 1143 GEP->insertBefore(Call); 1144 1145 // Generate a BitCast 1146 auto *BCInst2 = new BitCastInst(GEP, Call->getType()); 1147 BCInst2->insertBefore(Call); 1148 1149 // For the following code, 1150 // Block0: 1151 // ... 1152 // if (...) goto Block1 else ... 1153 // Block1: 1154 // %6 = load llvm.sk_buff:0:50$0:0:0:2:0 1155 // %7 = bitcast %struct.sk_buff* %2 to i8* 1156 // %8 = getelementptr i8, i8* %7, %6 1157 // ... 1158 // goto CommonExit 1159 // Block2: 1160 // ... 1161 // if (...) goto Block3 else ... 1162 // Block3: 1163 // %6 = load llvm.bpf_map:0:40$0:0:0:2:0 1164 // %7 = bitcast %struct.sk_buff* %2 to i8* 1165 // %8 = getelementptr i8, i8* %7, %6 1166 // ... 1167 // goto CommonExit 1168 // CommonExit 1169 // SimplifyCFG may generate: 1170 // Block0: 1171 // ... 1172 // if (...) goto Block_Common else ... 1173 // Block2: 1174 // ... 1175 // if (...) goto Block_Common else ... 1176 // Block_Common: 1177 // PHI = [llvm.sk_buff:0:50$0:0:0:2:0, llvm.bpf_map:0:40$0:0:0:2:0] 1178 // %6 = load PHI 1179 // %7 = bitcast %struct.sk_buff* %2 to i8* 1180 // %8 = getelementptr i8, i8* %7, %6 1181 // ... 1182 // goto CommonExit 1183 // For the above code, we cannot perform proper relocation since 1184 // "load PHI" has two possible relocations. 1185 // 1186 // To prevent above tail merging, we use __builtin_bpf_passthrough() 1187 // where one of its parameters is a seq_num. Since two 1188 // __builtin_bpf_passthrough() funcs will always have different seq_num, 1189 // tail merging cannot happen. The __builtin_bpf_passthrough() will be 1190 // removed in the beginning of Target IR passes. 1191 // 1192 // This approach is also used in other places when global var 1193 // representing a relocation is used. 1194 Instruction *PassThroughInst = 1195 BPFCoreSharedInfo::insertPassThrough(M, BB, BCInst2, Call); 1196 Call->replaceAllUsesWith(PassThroughInst); 1197 Call->eraseFromParent(); 1198 1199 return true; 1200 } 1201 1202 bool BPFAbstractMemberAccess::doTransformation(Function &F) { 1203 bool Transformed = false; 1204 1205 // Collect PreserveDIAccessIndex Intrinsic call chains. 1206 // The call chains will be used to generate the access 1207 // patterns similar to GEP. 1208 collectAICallChains(F); 1209 1210 for (auto &C : BaseAICalls) 1211 Transformed = transformGEPChain(C.first, C.second) || Transformed; 1212 1213 return removePreserveAccessIndexIntrinsic(F) || Transformed; 1214 } 1215 1216 PreservedAnalyses 1217 BPFAbstractMemberAccessPass::run(Function &F, FunctionAnalysisManager &AM) { 1218 return BPFAbstractMemberAccess(TM).run(F) ? PreservedAnalyses::none() 1219 : PreservedAnalyses::all(); 1220 } 1221