1 //===- WholeProgramDevirt.cpp - Whole program virtual call optimization ---===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass implements whole program optimization of virtual calls in cases
11 // where we know (via !type metadata) that the list of callees is fixed. This
12 // includes the following:
13 // - Single implementation devirtualization: if a virtual call has a single
14 // possible callee, replace all calls with a direct call to that callee.
15 // - Virtual constant propagation: if the virtual function's return type is an
16 // integer <=64 bits and all possible callees are readnone, for each class and
17 // each list of constant arguments: evaluate the function, store the return
18 // value alongside the virtual table, and rewrite each virtual call as a load
19 // from the virtual table.
20 // - Uniform return value optimization: if the conditions for virtual constant
21 // propagation hold and each function returns the same constant value, replace
22 // each virtual call with that constant.
23 // - Unique return value optimization for i1 return values: if the conditions
24 // for virtual constant propagation hold and a single vtable's function
25 // returns 0, or a single vtable's function returns 1, replace each virtual
26 // call with a comparison of the vptr against that vtable's address.
28 // This pass is intended to be used during the regular and thin LTO pipelines.
29 // During regular LTO, the pass determines the best optimization for each
30 // virtual call and applies the resolutions directly to virtual calls that are
31 // eligible for virtual call optimization (i.e. calls that use either of the
32 // llvm.assume(llvm.type.test) or llvm.type.checked.load intrinsics). During
33 // ThinLTO, the pass operates in two phases:
34 // - Export phase: this is run during the thin link over a single merged module
35 // that contains all vtables with !type metadata that participate in the link.
36 // The pass computes a resolution for each virtual call and stores it in the
37 // type identifier summary.
38 // - Import phase: this is run during the thin backends over the individual
39 // modules. The pass applies the resolutions previously computed during the
40 // import phase to each eligible virtual call.
42 //===----------------------------------------------------------------------===//
44 #include "llvm/Transforms/IPO/WholeProgramDevirt.h"
45 #include "llvm/ADT/ArrayRef.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/DenseMapInfo.h"
48 #include "llvm/ADT/DenseSet.h"
49 #include "llvm/ADT/MapVector.h"
50 #include "llvm/ADT/SmallVector.h"
51 #include "llvm/ADT/iterator_range.h"
52 #include "llvm/Analysis/AliasAnalysis.h"
53 #include "llvm/Analysis/BasicAliasAnalysis.h"
54 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
55 #include "llvm/Analysis/TypeMetadataUtils.h"
56 #include "llvm/IR/CallSite.h"
57 #include "llvm/IR/Constants.h"
58 #include "llvm/IR/DataLayout.h"
59 #include "llvm/IR/DebugLoc.h"
60 #include "llvm/IR/DerivedTypes.h"
61 #include "llvm/IR/Dominators.h"
62 #include "llvm/IR/Function.h"
63 #include "llvm/IR/GlobalAlias.h"
64 #include "llvm/IR/GlobalVariable.h"
65 #include "llvm/IR/IRBuilder.h"
66 #include "llvm/IR/InstrTypes.h"
67 #include "llvm/IR/Instruction.h"
68 #include "llvm/IR/Instructions.h"
69 #include "llvm/IR/Intrinsics.h"
70 #include "llvm/IR/LLVMContext.h"
71 #include "llvm/IR/Metadata.h"
72 #include "llvm/IR/Module.h"
73 #include "llvm/IR/ModuleSummaryIndexYAML.h"
74 #include "llvm/Pass.h"
75 #include "llvm/PassRegistry.h"
76 #include "llvm/PassSupport.h"
77 #include "llvm/Support/Casting.h"
78 #include "llvm/Support/Error.h"
79 #include "llvm/Support/FileSystem.h"
80 #include "llvm/Support/MathExtras.h"
81 #include "llvm/Transforms/IPO.h"
82 #include "llvm/Transforms/IPO/FunctionAttrs.h"
83 #include "llvm/Transforms/Utils/Evaluator.h"
91 using namespace wholeprogramdevirt;
93 #define DEBUG_TYPE "wholeprogramdevirt"
95 static cl::opt<PassSummaryAction> ClSummaryAction(
96 "wholeprogramdevirt-summary-action",
97 cl::desc("What to do with the summary when running this pass"),
98 cl::values(clEnumValN(PassSummaryAction::None, "none", "Do nothing"),
99 clEnumValN(PassSummaryAction::Import, "import",
100 "Import typeid resolutions from summary and globals"),
101 clEnumValN(PassSummaryAction::Export, "export",
102 "Export typeid resolutions to summary and globals")),
105 static cl::opt<std::string> ClReadSummary(
106 "wholeprogramdevirt-read-summary",
107 cl::desc("Read summary from given YAML file before running pass"),
110 static cl::opt<std::string> ClWriteSummary(
111 "wholeprogramdevirt-write-summary",
112 cl::desc("Write summary to given YAML file after running pass"),
115 static cl::opt<unsigned>
116 ClThreshold("wholeprogramdevirt-branch-funnel-threshold", cl::Hidden,
117 cl::init(10), cl::ZeroOrMore,
118 cl::desc("Maximum number of call targets per "
119 "call site to enable branch funnels"));
121 // Find the minimum offset that we may store a value of size Size bits at. If
122 // IsAfter is set, look for an offset before the object, otherwise look for an
123 // offset after the object.
125 wholeprogramdevirt::findLowestOffset(ArrayRef<VirtualCallTarget> Targets,
126 bool IsAfter, uint64_t Size) {
127 // Find a minimum offset taking into account only vtable sizes.
128 uint64_t MinByte = 0;
129 for (const VirtualCallTarget &Target : Targets) {
131 MinByte = std::max(MinByte, Target.minAfterBytes());
133 MinByte = std::max(MinByte, Target.minBeforeBytes());
136 // Build a vector of arrays of bytes covering, for each target, a slice of the
137 // used region (see AccumBitVector::BytesUsed in
138 // llvm/Transforms/IPO/WholeProgramDevirt.h) starting at MinByte. Effectively,
139 // this aligns the used regions to start at MinByte.
141 // In this example, A, B and C are vtables, # is a byte already allocated for
142 // a virtual function pointer, AAAA... (etc.) are the used regions for the
143 // vtables and Offset(X) is the value computed for the Offset variable below
149 // A: ################AAAAAAAA|AAAAAAAA
150 // B: ########BBBBBBBBBBBBBBBB|BBBB
151 // C: ########################|CCCCCCCCCCCCCCCC
154 // This code produces the slices of A, B and C that appear after the divider
156 std::vector<ArrayRef<uint8_t>> Used;
157 for (const VirtualCallTarget &Target : Targets) {
158 ArrayRef<uint8_t> VTUsed = IsAfter ? Target.TM->Bits->After.BytesUsed
159 : Target.TM->Bits->Before.BytesUsed;
160 uint64_t Offset = IsAfter ? MinByte - Target.minAfterBytes()
161 : MinByte - Target.minBeforeBytes();
163 // Disregard used regions that are smaller than Offset. These are
164 // effectively all-free regions that do not need to be checked.
165 if (VTUsed.size() > Offset)
166 Used.push_back(VTUsed.slice(Offset));
170 // Find a free bit in each member of Used.
171 for (unsigned I = 0;; ++I) {
172 uint8_t BitsUsed = 0;
173 for (auto &&B : Used)
176 if (BitsUsed != 0xff)
177 return (MinByte + I) * 8 +
178 countTrailingZeros(uint8_t(~BitsUsed), ZB_Undefined);
181 // Find a free (Size/8) byte region in each member of Used.
182 // FIXME: see if alignment helps.
183 for (unsigned I = 0;; ++I) {
184 for (auto &&B : Used) {
186 while ((I + Byte) < B.size() && Byte < (Size / 8)) {
192 return (MinByte + I) * 8;
198 void wholeprogramdevirt::setBeforeReturnValues(
199 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocBefore,
200 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) {
202 OffsetByte = -(AllocBefore / 8 + 1);
204 OffsetByte = -((AllocBefore + 7) / 8 + (BitWidth + 7) / 8);
205 OffsetBit = AllocBefore % 8;
207 for (VirtualCallTarget &Target : Targets) {
209 Target.setBeforeBit(AllocBefore);
211 Target.setBeforeBytes(AllocBefore, (BitWidth + 7) / 8);
215 void wholeprogramdevirt::setAfterReturnValues(
216 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocAfter,
217 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) {
219 OffsetByte = AllocAfter / 8;
221 OffsetByte = (AllocAfter + 7) / 8;
222 OffsetBit = AllocAfter % 8;
224 for (VirtualCallTarget &Target : Targets) {
226 Target.setAfterBit(AllocAfter);
228 Target.setAfterBytes(AllocAfter, (BitWidth + 7) / 8);
232 VirtualCallTarget::VirtualCallTarget(Function *Fn, const TypeMemberInfo *TM)
234 IsBigEndian(Fn->getParent()->getDataLayout().isBigEndian()), WasDevirt(false) {}
238 // A slot in a set of virtual tables. The TypeID identifies the set of virtual
239 // tables, and the ByteOffset is the offset in bytes from the address point to
240 // the virtual function pointer.
246 } // end anonymous namespace
250 template <> struct DenseMapInfo<VTableSlot> {
251 static VTableSlot getEmptyKey() {
252 return {DenseMapInfo<Metadata *>::getEmptyKey(),
253 DenseMapInfo<uint64_t>::getEmptyKey()};
255 static VTableSlot getTombstoneKey() {
256 return {DenseMapInfo<Metadata *>::getTombstoneKey(),
257 DenseMapInfo<uint64_t>::getTombstoneKey()};
259 static unsigned getHashValue(const VTableSlot &I) {
260 return DenseMapInfo<Metadata *>::getHashValue(I.TypeID) ^
261 DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset);
263 static bool isEqual(const VTableSlot &LHS,
264 const VTableSlot &RHS) {
265 return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset;
269 } // end namespace llvm
273 // A virtual call site. VTable is the loaded virtual table pointer, and CS is
274 // the indirect virtual call.
275 struct VirtualCallSite {
279 // If non-null, this field points to the associated unsafe use count stored in
280 // the DevirtModule::NumUnsafeUsesForTypeTest map below. See the description
281 // of that field for details.
282 unsigned *NumUnsafeUses;
285 emitRemark(const StringRef OptName, const StringRef TargetName,
286 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter) {
287 Function *F = CS.getCaller();
288 DebugLoc DLoc = CS->getDebugLoc();
289 BasicBlock *Block = CS.getParent();
292 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, OptName, DLoc, Block)
293 << NV("Optimization", OptName)
294 << ": devirtualized a call to "
295 << NV("FunctionName", TargetName));
298 void replaceAndErase(
299 const StringRef OptName, const StringRef TargetName, bool RemarksEnabled,
300 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter,
303 emitRemark(OptName, TargetName, OREGetter);
304 CS->replaceAllUsesWith(New);
305 if (auto II = dyn_cast<InvokeInst>(CS.getInstruction())) {
306 BranchInst::Create(II->getNormalDest(), CS.getInstruction());
307 II->getUnwindDest()->removePredecessor(II->getParent());
309 CS->eraseFromParent();
310 // This use is no longer unsafe.
316 // Call site information collected for a specific VTableSlot and possibly a list
317 // of constant integer arguments. The grouping by arguments is handled by the
318 // VTableSlotInfo class.
319 struct CallSiteInfo {
320 /// The set of call sites for this slot. Used during regular LTO and the
321 /// import phase of ThinLTO (as well as the export phase of ThinLTO for any
322 /// call sites that appear in the merged module itself); in each of these
323 /// cases we are directly operating on the call sites at the IR level.
324 std::vector<VirtualCallSite> CallSites;
326 /// Whether all call sites represented by this CallSiteInfo, including those
327 /// in summaries, have been devirtualized. This starts off as true because a
328 /// default constructed CallSiteInfo represents no call sites.
329 bool AllCallSitesDevirted = true;
331 // These fields are used during the export phase of ThinLTO and reflect
332 // information collected from function summaries.
334 /// Whether any function summary contains an llvm.assume(llvm.type.test) for
336 bool SummaryHasTypeTestAssumeUsers = false;
338 /// CFI-specific: a vector containing the list of function summaries that use
339 /// the llvm.type.checked.load intrinsic and therefore will require
340 /// resolutions for llvm.type.test in order to implement CFI checks if
341 /// devirtualization was unsuccessful. If devirtualization was successful, the
342 /// pass will clear this vector by calling markDevirt(). If at the end of the
343 /// pass the vector is non-empty, we will need to add a use of llvm.type.test
344 /// to each of the function summaries in the vector.
345 std::vector<FunctionSummary *> SummaryTypeCheckedLoadUsers;
347 bool isExported() const {
348 return SummaryHasTypeTestAssumeUsers ||
349 !SummaryTypeCheckedLoadUsers.empty();
352 void markSummaryHasTypeTestAssumeUsers() {
353 SummaryHasTypeTestAssumeUsers = true;
354 AllCallSitesDevirted = false;
357 void addSummaryTypeCheckedLoadUser(FunctionSummary *FS) {
358 SummaryTypeCheckedLoadUsers.push_back(FS);
359 AllCallSitesDevirted = false;
363 AllCallSitesDevirted = true;
365 // As explained in the comment for SummaryTypeCheckedLoadUsers.
366 SummaryTypeCheckedLoadUsers.clear();
370 // Call site information collected for a specific VTableSlot.
371 struct VTableSlotInfo {
372 // The set of call sites which do not have all constant integer arguments
373 // (excluding "this").
376 // The set of call sites with all constant integer arguments (excluding
377 // "this"), grouped by argument list.
378 std::map<std::vector<uint64_t>, CallSiteInfo> ConstCSInfo;
380 void addCallSite(Value *VTable, CallSite CS, unsigned *NumUnsafeUses);
383 CallSiteInfo &findCallSiteInfo(CallSite CS);
386 CallSiteInfo &VTableSlotInfo::findCallSiteInfo(CallSite CS) {
387 std::vector<uint64_t> Args;
388 auto *CI = dyn_cast<IntegerType>(CS.getType());
389 if (!CI || CI->getBitWidth() > 64 || CS.arg_empty())
391 for (auto &&Arg : make_range(CS.arg_begin() + 1, CS.arg_end())) {
392 auto *CI = dyn_cast<ConstantInt>(Arg);
393 if (!CI || CI->getBitWidth() > 64)
395 Args.push_back(CI->getZExtValue());
397 return ConstCSInfo[Args];
400 void VTableSlotInfo::addCallSite(Value *VTable, CallSite CS,
401 unsigned *NumUnsafeUses) {
402 auto &CSI = findCallSiteInfo(CS);
403 CSI.AllCallSitesDevirted = false;
404 CSI.CallSites.push_back({VTable, CS, NumUnsafeUses});
407 struct DevirtModule {
409 function_ref<AAResults &(Function &)> AARGetter;
410 function_ref<DominatorTree &(Function &)> LookupDomTree;
412 ModuleSummaryIndex *ExportSummary;
413 const ModuleSummaryIndex *ImportSummary;
416 PointerType *Int8PtrTy;
417 IntegerType *Int32Ty;
418 IntegerType *Int64Ty;
419 IntegerType *IntPtrTy;
422 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter;
424 MapVector<VTableSlot, VTableSlotInfo> CallSlots;
426 // This map keeps track of the number of "unsafe" uses of a loaded function
427 // pointer. The key is the associated llvm.type.test intrinsic call generated
428 // by this pass. An unsafe use is one that calls the loaded function pointer
429 // directly. Every time we eliminate an unsafe use (for example, by
430 // devirtualizing it or by applying virtual constant propagation), we
431 // decrement the value stored in this map. If a value reaches zero, we can
432 // eliminate the type check by RAUWing the associated llvm.type.test call with
434 std::map<CallInst *, unsigned> NumUnsafeUsesForTypeTest;
436 DevirtModule(Module &M, function_ref<AAResults &(Function &)> AARGetter,
437 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter,
438 function_ref<DominatorTree &(Function &)> LookupDomTree,
439 ModuleSummaryIndex *ExportSummary,
440 const ModuleSummaryIndex *ImportSummary)
441 : M(M), AARGetter(AARGetter), LookupDomTree(LookupDomTree),
442 ExportSummary(ExportSummary), ImportSummary(ImportSummary),
443 Int8Ty(Type::getInt8Ty(M.getContext())),
444 Int8PtrTy(Type::getInt8PtrTy(M.getContext())),
445 Int32Ty(Type::getInt32Ty(M.getContext())),
446 Int64Ty(Type::getInt64Ty(M.getContext())),
447 IntPtrTy(M.getDataLayout().getIntPtrType(M.getContext(), 0)),
448 RemarksEnabled(areRemarksEnabled()), OREGetter(OREGetter) {
449 assert(!(ExportSummary && ImportSummary));
452 bool areRemarksEnabled();
454 void scanTypeTestUsers(Function *TypeTestFunc, Function *AssumeFunc);
455 void scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc);
457 void buildTypeIdentifierMap(
458 std::vector<VTableBits> &Bits,
459 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap);
460 Constant *getPointerAtOffset(Constant *I, uint64_t Offset);
462 tryFindVirtualCallTargets(std::vector<VirtualCallTarget> &TargetsForSlot,
463 const std::set<TypeMemberInfo> &TypeMemberInfos,
464 uint64_t ByteOffset);
466 void applySingleImplDevirt(VTableSlotInfo &SlotInfo, Constant *TheFn,
468 bool trySingleImplDevirt(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
469 VTableSlotInfo &SlotInfo,
470 WholeProgramDevirtResolution *Res);
472 void applyICallBranchFunnel(VTableSlotInfo &SlotInfo, Constant *JT,
474 void tryICallBranchFunnel(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
475 VTableSlotInfo &SlotInfo,
476 WholeProgramDevirtResolution *Res, VTableSlot Slot);
478 bool tryEvaluateFunctionsWithArgs(
479 MutableArrayRef<VirtualCallTarget> TargetsForSlot,
480 ArrayRef<uint64_t> Args);
482 void applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
484 bool tryUniformRetValOpt(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
485 CallSiteInfo &CSInfo,
486 WholeProgramDevirtResolution::ByArg *Res);
488 // Returns the global symbol name that is used to export information about the
489 // given vtable slot and list of arguments.
490 std::string getGlobalName(VTableSlot Slot, ArrayRef<uint64_t> Args,
493 bool shouldExportConstantsAsAbsoluteSymbols();
495 // This function is called during the export phase to create a symbol
496 // definition containing information about the given vtable slot and list of
498 void exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name,
500 void exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name,
501 uint32_t Const, uint32_t &Storage);
503 // This function is called during the import phase to create a reference to
504 // the symbol definition created during the export phase.
505 Constant *importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
507 Constant *importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args,
508 StringRef Name, IntegerType *IntTy,
511 Constant *getMemberAddr(const TypeMemberInfo *M);
513 void applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, bool IsOne,
514 Constant *UniqueMemberAddr);
515 bool tryUniqueRetValOpt(unsigned BitWidth,
516 MutableArrayRef<VirtualCallTarget> TargetsForSlot,
517 CallSiteInfo &CSInfo,
518 WholeProgramDevirtResolution::ByArg *Res,
519 VTableSlot Slot, ArrayRef<uint64_t> Args);
521 void applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName,
522 Constant *Byte, Constant *Bit);
523 bool tryVirtualConstProp(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
524 VTableSlotInfo &SlotInfo,
525 WholeProgramDevirtResolution *Res, VTableSlot Slot);
527 void rebuildGlobal(VTableBits &B);
529 // Apply the summary resolution for Slot to all virtual calls in SlotInfo.
530 void importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo);
532 // If we were able to eliminate all unsafe uses for a type checked load,
533 // eliminate the associated type tests by replacing them with true.
534 void removeRedundantTypeTests();
538 // Lower the module using the action and summary passed as command line
539 // arguments. For testing purposes only.
541 runForTesting(Module &M, function_ref<AAResults &(Function &)> AARGetter,
542 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter,
543 function_ref<DominatorTree &(Function &)> LookupDomTree);
546 struct WholeProgramDevirt : public ModulePass {
549 bool UseCommandLine = false;
551 ModuleSummaryIndex *ExportSummary;
552 const ModuleSummaryIndex *ImportSummary;
554 WholeProgramDevirt() : ModulePass(ID), UseCommandLine(true) {
555 initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry());
558 WholeProgramDevirt(ModuleSummaryIndex *ExportSummary,
559 const ModuleSummaryIndex *ImportSummary)
560 : ModulePass(ID), ExportSummary(ExportSummary),
561 ImportSummary(ImportSummary) {
562 initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry());
565 bool runOnModule(Module &M) override {
569 // In the new pass manager, we can request the optimization
570 // remark emitter pass on a per-function-basis, which the
571 // OREGetter will do for us.
572 // In the old pass manager, this is harder, so we just build
573 // an optimization remark emitter on the fly, when we need it.
574 std::unique_ptr<OptimizationRemarkEmitter> ORE;
575 auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & {
576 ORE = make_unique<OptimizationRemarkEmitter>(F);
580 auto LookupDomTree = [this](Function &F) -> DominatorTree & {
581 return this->getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
585 return DevirtModule::runForTesting(M, LegacyAARGetter(*this), OREGetter,
588 return DevirtModule(M, LegacyAARGetter(*this), OREGetter, LookupDomTree,
589 ExportSummary, ImportSummary)
593 void getAnalysisUsage(AnalysisUsage &AU) const override {
594 AU.addRequired<AssumptionCacheTracker>();
595 AU.addRequired<TargetLibraryInfoWrapperPass>();
596 AU.addRequired<DominatorTreeWrapperPass>();
600 } // end anonymous namespace
602 INITIALIZE_PASS_BEGIN(WholeProgramDevirt, "wholeprogramdevirt",
603 "Whole program devirtualization", false, false)
604 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
605 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
606 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
607 INITIALIZE_PASS_END(WholeProgramDevirt, "wholeprogramdevirt",
608 "Whole program devirtualization", false, false)
609 char WholeProgramDevirt::ID = 0;
612 llvm::createWholeProgramDevirtPass(ModuleSummaryIndex *ExportSummary,
613 const ModuleSummaryIndex *ImportSummary) {
614 return new WholeProgramDevirt(ExportSummary, ImportSummary);
617 PreservedAnalyses WholeProgramDevirtPass::run(Module &M,
618 ModuleAnalysisManager &AM) {
619 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
620 auto AARGetter = [&](Function &F) -> AAResults & {
621 return FAM.getResult<AAManager>(F);
623 auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & {
624 return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F);
626 auto LookupDomTree = [&FAM](Function &F) -> DominatorTree & {
627 return FAM.getResult<DominatorTreeAnalysis>(F);
629 if (!DevirtModule(M, AARGetter, OREGetter, LookupDomTree, ExportSummary,
632 return PreservedAnalyses::all();
633 return PreservedAnalyses::none();
636 bool DevirtModule::runForTesting(
637 Module &M, function_ref<AAResults &(Function &)> AARGetter,
638 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter,
639 function_ref<DominatorTree &(Function &)> LookupDomTree) {
640 ModuleSummaryIndex Summary(/*HaveGVs=*/false);
642 // Handle the command-line summary arguments. This code is for testing
643 // purposes only, so we handle errors directly.
644 if (!ClReadSummary.empty()) {
645 ExitOnError ExitOnErr("-wholeprogramdevirt-read-summary: " + ClReadSummary +
647 auto ReadSummaryFile =
648 ExitOnErr(errorOrToExpected(MemoryBuffer::getFile(ClReadSummary)));
650 yaml::Input In(ReadSummaryFile->getBuffer());
652 ExitOnErr(errorCodeToError(In.error()));
657 M, AARGetter, OREGetter, LookupDomTree,
658 ClSummaryAction == PassSummaryAction::Export ? &Summary : nullptr,
659 ClSummaryAction == PassSummaryAction::Import ? &Summary : nullptr)
662 if (!ClWriteSummary.empty()) {
663 ExitOnError ExitOnErr(
664 "-wholeprogramdevirt-write-summary: " + ClWriteSummary + ": ");
666 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::F_Text);
667 ExitOnErr(errorCodeToError(EC));
669 yaml::Output Out(OS);
676 void DevirtModule::buildTypeIdentifierMap(
677 std::vector<VTableBits> &Bits,
678 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) {
679 DenseMap<GlobalVariable *, VTableBits *> GVToBits;
680 Bits.reserve(M.getGlobalList().size());
681 SmallVector<MDNode *, 2> Types;
682 for (GlobalVariable &GV : M.globals()) {
684 GV.getMetadata(LLVMContext::MD_type, Types);
685 if (GV.isDeclaration() || Types.empty())
688 VTableBits *&BitsPtr = GVToBits[&GV];
691 Bits.back().GV = &GV;
692 Bits.back().ObjectSize =
693 M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType());
694 BitsPtr = &Bits.back();
697 for (MDNode *Type : Types) {
698 auto TypeID = Type->getOperand(1).get();
702 cast<ConstantAsMetadata>(Type->getOperand(0))->getValue())
705 TypeIdMap[TypeID].insert({BitsPtr, Offset});
710 Constant *DevirtModule::getPointerAtOffset(Constant *I, uint64_t Offset) {
711 if (I->getType()->isPointerTy()) {
717 const DataLayout &DL = M.getDataLayout();
719 if (auto *C = dyn_cast<ConstantStruct>(I)) {
720 const StructLayout *SL = DL.getStructLayout(C->getType());
721 if (Offset >= SL->getSizeInBytes())
724 unsigned Op = SL->getElementContainingOffset(Offset);
725 return getPointerAtOffset(cast<Constant>(I->getOperand(Op)),
726 Offset - SL->getElementOffset(Op));
728 if (auto *C = dyn_cast<ConstantArray>(I)) {
729 ArrayType *VTableTy = C->getType();
730 uint64_t ElemSize = DL.getTypeAllocSize(VTableTy->getElementType());
732 unsigned Op = Offset / ElemSize;
733 if (Op >= C->getNumOperands())
736 return getPointerAtOffset(cast<Constant>(I->getOperand(Op)),
742 bool DevirtModule::tryFindVirtualCallTargets(
743 std::vector<VirtualCallTarget> &TargetsForSlot,
744 const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset) {
745 for (const TypeMemberInfo &TM : TypeMemberInfos) {
746 if (!TM.Bits->GV->isConstant())
749 Constant *Ptr = getPointerAtOffset(TM.Bits->GV->getInitializer(),
750 TM.Offset + ByteOffset);
754 auto Fn = dyn_cast<Function>(Ptr->stripPointerCasts());
758 // We can disregard __cxa_pure_virtual as a possible call target, as
759 // calls to pure virtuals are UB.
760 if (Fn->getName() == "__cxa_pure_virtual")
763 TargetsForSlot.push_back({Fn, &TM});
766 // Give up if we couldn't find any targets.
767 return !TargetsForSlot.empty();
770 void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo,
771 Constant *TheFn, bool &IsExported) {
772 auto Apply = [&](CallSiteInfo &CSInfo) {
773 for (auto &&VCallSite : CSInfo.CallSites) {
775 VCallSite.emitRemark("single-impl",
776 TheFn->stripPointerCasts()->getName(), OREGetter);
777 VCallSite.CS.setCalledFunction(ConstantExpr::getBitCast(
778 TheFn, VCallSite.CS.getCalledValue()->getType()));
779 // This use is no longer unsafe.
780 if (VCallSite.NumUnsafeUses)
781 --*VCallSite.NumUnsafeUses;
783 if (CSInfo.isExported())
787 Apply(SlotInfo.CSInfo);
788 for (auto &P : SlotInfo.ConstCSInfo)
792 bool DevirtModule::trySingleImplDevirt(
793 MutableArrayRef<VirtualCallTarget> TargetsForSlot,
794 VTableSlotInfo &SlotInfo, WholeProgramDevirtResolution *Res) {
795 // See if the program contains a single implementation of this virtual
797 Function *TheFn = TargetsForSlot[0].Fn;
798 for (auto &&Target : TargetsForSlot)
799 if (TheFn != Target.Fn)
802 // If so, update each call site to call that implementation directly.
804 TargetsForSlot[0].WasDevirt = true;
806 bool IsExported = false;
807 applySingleImplDevirt(SlotInfo, TheFn, IsExported);
811 // If the only implementation has local linkage, we must promote to external
812 // to make it visible to thin LTO objects. We can only get here during the
813 // ThinLTO export phase.
814 if (TheFn->hasLocalLinkage()) {
815 std::string NewName = (TheFn->getName() + "$merged").str();
817 // Since we are renaming the function, any comdats with the same name must
818 // also be renamed. This is required when targeting COFF, as the comdat name
819 // must match one of the names of the symbols in the comdat.
820 if (Comdat *C = TheFn->getComdat()) {
821 if (C->getName() == TheFn->getName()) {
822 Comdat *NewC = M.getOrInsertComdat(NewName);
823 NewC->setSelectionKind(C->getSelectionKind());
824 for (GlobalObject &GO : M.global_objects())
825 if (GO.getComdat() == C)
830 TheFn->setLinkage(GlobalValue::ExternalLinkage);
831 TheFn->setVisibility(GlobalValue::HiddenVisibility);
832 TheFn->setName(NewName);
835 Res->TheKind = WholeProgramDevirtResolution::SingleImpl;
836 Res->SingleImplName = TheFn->getName();
841 void DevirtModule::tryICallBranchFunnel(
842 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo,
843 WholeProgramDevirtResolution *Res, VTableSlot Slot) {
844 Triple T(M.getTargetTriple());
845 if (T.getArch() != Triple::x86_64)
848 if (TargetsForSlot.size() > ClThreshold)
851 bool HasNonDevirt = !SlotInfo.CSInfo.AllCallSitesDevirted;
853 for (auto &P : SlotInfo.ConstCSInfo)
854 if (!P.second.AllCallSitesDevirted) {
863 FunctionType::get(Type::getVoidTy(M.getContext()), {Int8PtrTy}, true);
865 if (isa<MDString>(Slot.TypeID)) {
866 JT = Function::Create(FT, Function::ExternalLinkage,
867 M.getDataLayout().getProgramAddressSpace(),
868 getGlobalName(Slot, {}, "branch_funnel"), &M);
869 JT->setVisibility(GlobalValue::HiddenVisibility);
871 JT = Function::Create(FT, Function::InternalLinkage,
872 M.getDataLayout().getProgramAddressSpace(),
873 "branch_funnel", &M);
875 JT->addAttribute(1, Attribute::Nest);
877 std::vector<Value *> JTArgs;
878 JTArgs.push_back(JT->arg_begin());
879 for (auto &T : TargetsForSlot) {
880 JTArgs.push_back(getMemberAddr(T.TM));
881 JTArgs.push_back(T.Fn);
884 BasicBlock *BB = BasicBlock::Create(M.getContext(), "", JT, nullptr);
886 Intrinsic::getDeclaration(&M, llvm::Intrinsic::icall_branch_funnel, {});
888 auto *CI = CallInst::Create(Intr, JTArgs, "", BB);
889 CI->setTailCallKind(CallInst::TCK_MustTail);
890 ReturnInst::Create(M.getContext(), nullptr, BB);
892 bool IsExported = false;
893 applyICallBranchFunnel(SlotInfo, JT, IsExported);
895 Res->TheKind = WholeProgramDevirtResolution::BranchFunnel;
898 void DevirtModule::applyICallBranchFunnel(VTableSlotInfo &SlotInfo,
899 Constant *JT, bool &IsExported) {
900 auto Apply = [&](CallSiteInfo &CSInfo) {
901 if (CSInfo.isExported())
903 if (CSInfo.AllCallSitesDevirted)
905 for (auto &&VCallSite : CSInfo.CallSites) {
906 CallSite CS = VCallSite.CS;
908 // Jump tables are only profitable if the retpoline mitigation is enabled.
909 Attribute FSAttr = CS.getCaller()->getFnAttribute("target-features");
910 if (FSAttr.hasAttribute(Attribute::None) ||
911 !FSAttr.getValueAsString().contains("+retpoline"))
915 VCallSite.emitRemark("branch-funnel",
916 JT->stripPointerCasts()->getName(), OREGetter);
918 // Pass the address of the vtable in the nest register, which is r10 on
920 std::vector<Type *> NewArgs;
921 NewArgs.push_back(Int8PtrTy);
922 for (Type *T : CS.getFunctionType()->params())
923 NewArgs.push_back(T);
924 PointerType *NewFT = PointerType::getUnqual(
925 FunctionType::get(CS.getFunctionType()->getReturnType(), NewArgs,
926 CS.getFunctionType()->isVarArg()));
928 IRBuilder<> IRB(CS.getInstruction());
929 std::vector<Value *> Args;
930 Args.push_back(IRB.CreateBitCast(VCallSite.VTable, Int8PtrTy));
931 for (unsigned I = 0; I != CS.getNumArgOperands(); ++I)
932 Args.push_back(CS.getArgOperand(I));
936 NewCS = IRB.CreateCall(IRB.CreateBitCast(JT, NewFT), Args);
938 NewCS = IRB.CreateInvoke(
939 IRB.CreateBitCast(JT, NewFT),
940 cast<InvokeInst>(CS.getInstruction())->getNormalDest(),
941 cast<InvokeInst>(CS.getInstruction())->getUnwindDest(), Args);
942 NewCS.setCallingConv(CS.getCallingConv());
944 AttributeList Attrs = CS.getAttributes();
945 std::vector<AttributeSet> NewArgAttrs;
946 NewArgAttrs.push_back(AttributeSet::get(
947 M.getContext(), ArrayRef<Attribute>{Attribute::get(
948 M.getContext(), Attribute::Nest)}));
949 for (unsigned I = 0; I + 2 < Attrs.getNumAttrSets(); ++I)
950 NewArgAttrs.push_back(Attrs.getParamAttributes(I));
952 AttributeList::get(M.getContext(), Attrs.getFnAttributes(),
953 Attrs.getRetAttributes(), NewArgAttrs));
955 CS->replaceAllUsesWith(NewCS.getInstruction());
956 CS->eraseFromParent();
958 // This use is no longer unsafe.
959 if (VCallSite.NumUnsafeUses)
960 --*VCallSite.NumUnsafeUses;
962 // Don't mark as devirtualized because there may be callers compiled without
963 // retpoline mitigation, which would mean that they are lowered to
964 // llvm.type.test and therefore require an llvm.type.test resolution for the
967 Apply(SlotInfo.CSInfo);
968 for (auto &P : SlotInfo.ConstCSInfo)
972 bool DevirtModule::tryEvaluateFunctionsWithArgs(
973 MutableArrayRef<VirtualCallTarget> TargetsForSlot,
974 ArrayRef<uint64_t> Args) {
975 // Evaluate each function and store the result in each target's RetVal
977 for (VirtualCallTarget &Target : TargetsForSlot) {
978 if (Target.Fn->arg_size() != Args.size() + 1)
981 Evaluator Eval(M.getDataLayout(), nullptr);
982 SmallVector<Constant *, 2> EvalArgs;
984 Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0)));
985 for (unsigned I = 0; I != Args.size(); ++I) {
986 auto *ArgTy = dyn_cast<IntegerType>(
987 Target.Fn->getFunctionType()->getParamType(I + 1));
990 EvalArgs.push_back(ConstantInt::get(ArgTy, Args[I]));
994 if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) ||
995 !isa<ConstantInt>(RetVal))
997 Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue();
1002 void DevirtModule::applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
1003 uint64_t TheRetVal) {
1004 for (auto Call : CSInfo.CallSites)
1005 Call.replaceAndErase(
1006 "uniform-ret-val", FnName, RemarksEnabled, OREGetter,
1007 ConstantInt::get(cast<IntegerType>(Call.CS.getType()), TheRetVal));
1008 CSInfo.markDevirt();
1011 bool DevirtModule::tryUniformRetValOpt(
1012 MutableArrayRef<VirtualCallTarget> TargetsForSlot, CallSiteInfo &CSInfo,
1013 WholeProgramDevirtResolution::ByArg *Res) {
1014 // Uniform return value optimization. If all functions return the same
1015 // constant, replace all calls with that constant.
1016 uint64_t TheRetVal = TargetsForSlot[0].RetVal;
1017 for (const VirtualCallTarget &Target : TargetsForSlot)
1018 if (Target.RetVal != TheRetVal)
1021 if (CSInfo.isExported()) {
1022 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniformRetVal;
1023 Res->Info = TheRetVal;
1026 applyUniformRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), TheRetVal);
1028 for (auto &&Target : TargetsForSlot)
1029 Target.WasDevirt = true;
1033 std::string DevirtModule::getGlobalName(VTableSlot Slot,
1034 ArrayRef<uint64_t> Args,
1036 std::string FullName = "__typeid_";
1037 raw_string_ostream OS(FullName);
1038 OS << cast<MDString>(Slot.TypeID)->getString() << '_' << Slot.ByteOffset;
1039 for (uint64_t Arg : Args)
1045 bool DevirtModule::shouldExportConstantsAsAbsoluteSymbols() {
1046 Triple T(M.getTargetTriple());
1047 return (T.getArch() == Triple::x86 || T.getArch() == Triple::x86_64) &&
1048 T.getObjectFormat() == Triple::ELF;
1051 void DevirtModule::exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
1052 StringRef Name, Constant *C) {
1053 GlobalAlias *GA = GlobalAlias::create(Int8Ty, 0, GlobalValue::ExternalLinkage,
1054 getGlobalName(Slot, Args, Name), C, &M);
1055 GA->setVisibility(GlobalValue::HiddenVisibility);
1058 void DevirtModule::exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args,
1059 StringRef Name, uint32_t Const,
1060 uint32_t &Storage) {
1061 if (shouldExportConstantsAsAbsoluteSymbols()) {
1064 ConstantExpr::getIntToPtr(ConstantInt::get(Int32Ty, Const), Int8PtrTy));
1071 Constant *DevirtModule::importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
1073 Constant *C = M.getOrInsertGlobal(getGlobalName(Slot, Args, Name), Int8Ty);
1074 auto *GV = dyn_cast<GlobalVariable>(C);
1076 GV->setVisibility(GlobalValue::HiddenVisibility);
1080 Constant *DevirtModule::importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args,
1081 StringRef Name, IntegerType *IntTy,
1083 if (!shouldExportConstantsAsAbsoluteSymbols())
1084 return ConstantInt::get(IntTy, Storage);
1086 Constant *C = importGlobal(Slot, Args, Name);
1087 auto *GV = cast<GlobalVariable>(C->stripPointerCasts());
1088 C = ConstantExpr::getPtrToInt(C, IntTy);
1090 // We only need to set metadata if the global is newly created, in which
1091 // case it would not have hidden visibility.
1092 if (GV->hasMetadata(LLVMContext::MD_absolute_symbol))
1095 auto SetAbsRange = [&](uint64_t Min, uint64_t Max) {
1096 auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Min));
1097 auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Max));
1098 GV->setMetadata(LLVMContext::MD_absolute_symbol,
1099 MDNode::get(M.getContext(), {MinC, MaxC}));
1101 unsigned AbsWidth = IntTy->getBitWidth();
1102 if (AbsWidth == IntPtrTy->getBitWidth())
1103 SetAbsRange(~0ull, ~0ull); // Full set.
1105 SetAbsRange(0, 1ull << AbsWidth);
1109 void DevirtModule::applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
1111 Constant *UniqueMemberAddr) {
1112 for (auto &&Call : CSInfo.CallSites) {
1113 IRBuilder<> B(Call.CS.getInstruction());
1115 B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
1116 B.CreateBitCast(Call.VTable, Int8PtrTy), UniqueMemberAddr);
1117 Cmp = B.CreateZExt(Cmp, Call.CS->getType());
1118 Call.replaceAndErase("unique-ret-val", FnName, RemarksEnabled, OREGetter,
1121 CSInfo.markDevirt();
1124 Constant *DevirtModule::getMemberAddr(const TypeMemberInfo *M) {
1125 Constant *C = ConstantExpr::getBitCast(M->Bits->GV, Int8PtrTy);
1126 return ConstantExpr::getGetElementPtr(Int8Ty, C,
1127 ConstantInt::get(Int64Ty, M->Offset));
1130 bool DevirtModule::tryUniqueRetValOpt(
1131 unsigned BitWidth, MutableArrayRef<VirtualCallTarget> TargetsForSlot,
1132 CallSiteInfo &CSInfo, WholeProgramDevirtResolution::ByArg *Res,
1133 VTableSlot Slot, ArrayRef<uint64_t> Args) {
1134 // IsOne controls whether we look for a 0 or a 1.
1135 auto tryUniqueRetValOptFor = [&](bool IsOne) {
1136 const TypeMemberInfo *UniqueMember = nullptr;
1137 for (const VirtualCallTarget &Target : TargetsForSlot) {
1138 if (Target.RetVal == (IsOne ? 1 : 0)) {
1141 UniqueMember = Target.TM;
1145 // We should have found a unique member or bailed out by now. We already
1146 // checked for a uniform return value in tryUniformRetValOpt.
1147 assert(UniqueMember);
1149 Constant *UniqueMemberAddr = getMemberAddr(UniqueMember);
1150 if (CSInfo.isExported()) {
1151 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniqueRetVal;
1154 exportGlobal(Slot, Args, "unique_member", UniqueMemberAddr);
1157 // Replace each call with the comparison.
1158 applyUniqueRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), IsOne,
1161 // Update devirtualization statistics for targets.
1163 for (auto &&Target : TargetsForSlot)
1164 Target.WasDevirt = true;
1169 if (BitWidth == 1) {
1170 if (tryUniqueRetValOptFor(true))
1172 if (tryUniqueRetValOptFor(false))
1178 void DevirtModule::applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName,
1179 Constant *Byte, Constant *Bit) {
1180 for (auto Call : CSInfo.CallSites) {
1181 auto *RetType = cast<IntegerType>(Call.CS.getType());
1182 IRBuilder<> B(Call.CS.getInstruction());
1184 B.CreateGEP(Int8Ty, B.CreateBitCast(Call.VTable, Int8PtrTy), Byte);
1185 if (RetType->getBitWidth() == 1) {
1186 Value *Bits = B.CreateLoad(Addr);
1187 Value *BitsAndBit = B.CreateAnd(Bits, Bit);
1188 auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0));
1189 Call.replaceAndErase("virtual-const-prop-1-bit", FnName, RemarksEnabled,
1190 OREGetter, IsBitSet);
1192 Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo());
1193 Value *Val = B.CreateLoad(RetType, ValAddr);
1194 Call.replaceAndErase("virtual-const-prop", FnName, RemarksEnabled,
1198 CSInfo.markDevirt();
1201 bool DevirtModule::tryVirtualConstProp(
1202 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo,
1203 WholeProgramDevirtResolution *Res, VTableSlot Slot) {
1204 // This only works if the function returns an integer.
1205 auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType());
1208 unsigned BitWidth = RetType->getBitWidth();
1212 // Make sure that each function is defined, does not access memory, takes at
1213 // least one argument, does not use its first argument (which we assume is
1214 // 'this'), and has the same return type.
1216 // Note that we test whether this copy of the function is readnone, rather
1217 // than testing function attributes, which must hold for any copy of the
1218 // function, even a less optimized version substituted at link time. This is
1219 // sound because the virtual constant propagation optimizations effectively
1220 // inline all implementations of the virtual function into each call site,
1221 // rather than using function attributes to perform local optimization.
1222 for (VirtualCallTarget &Target : TargetsForSlot) {
1223 if (Target.Fn->isDeclaration() ||
1224 computeFunctionBodyMemoryAccess(*Target.Fn, AARGetter(*Target.Fn)) !=
1226 Target.Fn->arg_empty() || !Target.Fn->arg_begin()->use_empty() ||
1227 Target.Fn->getReturnType() != RetType)
1231 for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) {
1232 if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first))
1235 WholeProgramDevirtResolution::ByArg *ResByArg = nullptr;
1237 ResByArg = &Res->ResByArg[CSByConstantArg.first];
1239 if (tryUniformRetValOpt(TargetsForSlot, CSByConstantArg.second, ResByArg))
1242 if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second,
1243 ResByArg, Slot, CSByConstantArg.first))
1246 // Find an allocation offset in bits in all vtables associated with the
1248 uint64_t AllocBefore =
1249 findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth);
1250 uint64_t AllocAfter =
1251 findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth);
1253 // Calculate the total amount of padding needed to store a value at both
1254 // ends of the object.
1255 uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0;
1256 for (auto &&Target : TargetsForSlot) {
1257 TotalPaddingBefore += std::max<int64_t>(
1258 (AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0);
1259 TotalPaddingAfter += std::max<int64_t>(
1260 (AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0);
1263 // If the amount of padding is too large, give up.
1264 // FIXME: do something smarter here.
1265 if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128)
1268 // Calculate the offset to the value as a (possibly negative) byte offset
1269 // and (if applicable) a bit offset, and store the values in the targets.
1272 if (TotalPaddingBefore <= TotalPaddingAfter)
1273 setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte,
1276 setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte,
1280 for (auto &&Target : TargetsForSlot)
1281 Target.WasDevirt = true;
1284 if (CSByConstantArg.second.isExported()) {
1285 ResByArg->TheKind = WholeProgramDevirtResolution::ByArg::VirtualConstProp;
1286 exportConstant(Slot, CSByConstantArg.first, "byte", OffsetByte,
1288 exportConstant(Slot, CSByConstantArg.first, "bit", 1ULL << OffsetBit,
1292 // Rewrite each call to a load from OffsetByte/OffsetBit.
1293 Constant *ByteConst = ConstantInt::get(Int32Ty, OffsetByte);
1294 Constant *BitConst = ConstantInt::get(Int8Ty, 1ULL << OffsetBit);
1295 applyVirtualConstProp(CSByConstantArg.second,
1296 TargetsForSlot[0].Fn->getName(), ByteConst, BitConst);
1301 void DevirtModule::rebuildGlobal(VTableBits &B) {
1302 if (B.Before.Bytes.empty() && B.After.Bytes.empty())
1305 // Align each byte array to pointer width.
1306 unsigned PointerSize = M.getDataLayout().getPointerSize();
1307 B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), PointerSize));
1308 B.After.Bytes.resize(alignTo(B.After.Bytes.size(), PointerSize));
1310 // Before was stored in reverse order; flip it now.
1311 for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I)
1312 std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]);
1314 // Build an anonymous global containing the before bytes, followed by the
1315 // original initializer, followed by the after bytes.
1316 auto NewInit = ConstantStruct::getAnon(
1317 {ConstantDataArray::get(M.getContext(), B.Before.Bytes),
1318 B.GV->getInitializer(),
1319 ConstantDataArray::get(M.getContext(), B.After.Bytes)});
1321 new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(),
1322 GlobalVariable::PrivateLinkage, NewInit, "", B.GV);
1323 NewGV->setSection(B.GV->getSection());
1324 NewGV->setComdat(B.GV->getComdat());
1326 // Copy the original vtable's metadata to the anonymous global, adjusting
1327 // offsets as required.
1328 NewGV->copyMetadata(B.GV, B.Before.Bytes.size());
1330 // Build an alias named after the original global, pointing at the second
1331 // element (the original initializer).
1332 auto Alias = GlobalAlias::create(
1333 B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "",
1334 ConstantExpr::getGetElementPtr(
1335 NewInit->getType(), NewGV,
1336 ArrayRef<Constant *>{ConstantInt::get(Int32Ty, 0),
1337 ConstantInt::get(Int32Ty, 1)}),
1339 Alias->setVisibility(B.GV->getVisibility());
1340 Alias->takeName(B.GV);
1342 B.GV->replaceAllUsesWith(Alias);
1343 B.GV->eraseFromParent();
1346 bool DevirtModule::areRemarksEnabled() {
1347 const auto &FL = M.getFunctionList();
1348 for (const Function &Fn : FL) {
1349 const auto &BBL = Fn.getBasicBlockList();
1352 auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBL.front());
1353 return DI.isEnabled();
1358 void DevirtModule::scanTypeTestUsers(Function *TypeTestFunc,
1359 Function *AssumeFunc) {
1360 // Find all virtual calls via a virtual table pointer %p under an assumption
1361 // of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p
1362 // points to a member of the type identifier %md. Group calls by (type ID,
1363 // offset) pair (effectively the identity of the virtual function) and store
1365 DenseSet<CallSite> SeenCallSites;
1366 for (auto I = TypeTestFunc->use_begin(), E = TypeTestFunc->use_end();
1368 auto CI = dyn_cast<CallInst>(I->getUser());
1373 // Search for virtual calls based on %p and add them to DevirtCalls.
1374 SmallVector<DevirtCallSite, 1> DevirtCalls;
1375 SmallVector<CallInst *, 1> Assumes;
1376 auto &DT = LookupDomTree(*CI->getFunction());
1377 findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT);
1379 // If we found any, add them to CallSlots.
1380 if (!Assumes.empty()) {
1382 cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata();
1383 Value *Ptr = CI->getArgOperand(0)->stripPointerCasts();
1384 for (DevirtCallSite Call : DevirtCalls) {
1385 // Only add this CallSite if we haven't seen it before. The vtable
1386 // pointer may have been CSE'd with pointers from other call sites,
1387 // and we don't want to process call sites multiple times. We can't
1388 // just skip the vtable Ptr if it has been seen before, however, since
1389 // it may be shared by type tests that dominate different calls.
1390 if (SeenCallSites.insert(Call.CS).second)
1391 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS, nullptr);
1395 // We no longer need the assumes or the type test.
1396 for (auto Assume : Assumes)
1397 Assume->eraseFromParent();
1398 // We can't use RecursivelyDeleteTriviallyDeadInstructions here because we
1399 // may use the vtable argument later.
1400 if (CI->use_empty())
1401 CI->eraseFromParent();
1405 void DevirtModule::scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc) {
1406 Function *TypeTestFunc = Intrinsic::getDeclaration(&M, Intrinsic::type_test);
1408 for (auto I = TypeCheckedLoadFunc->use_begin(),
1409 E = TypeCheckedLoadFunc->use_end();
1411 auto CI = dyn_cast<CallInst>(I->getUser());
1416 Value *Ptr = CI->getArgOperand(0);
1417 Value *Offset = CI->getArgOperand(1);
1418 Value *TypeIdValue = CI->getArgOperand(2);
1419 Metadata *TypeId = cast<MetadataAsValue>(TypeIdValue)->getMetadata();
1421 SmallVector<DevirtCallSite, 1> DevirtCalls;
1422 SmallVector<Instruction *, 1> LoadedPtrs;
1423 SmallVector<Instruction *, 1> Preds;
1424 bool HasNonCallUses = false;
1425 auto &DT = LookupDomTree(*CI->getFunction());
1426 findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds,
1427 HasNonCallUses, CI, DT);
1429 // Start by generating "pessimistic" code that explicitly loads the function
1430 // pointer from the vtable and performs the type check. If possible, we will
1431 // eliminate the load and the type check later.
1433 // If possible, only generate the load at the point where it is used.
1434 // This helps avoid unnecessary spills.
1436 (LoadedPtrs.size() == 1 && !HasNonCallUses) ? LoadedPtrs[0] : CI);
1437 Value *GEP = LoadB.CreateGEP(Int8Ty, Ptr, Offset);
1438 Value *GEPPtr = LoadB.CreateBitCast(GEP, PointerType::getUnqual(Int8PtrTy));
1439 Value *LoadedValue = LoadB.CreateLoad(Int8PtrTy, GEPPtr);
1441 for (Instruction *LoadedPtr : LoadedPtrs) {
1442 LoadedPtr->replaceAllUsesWith(LoadedValue);
1443 LoadedPtr->eraseFromParent();
1446 // Likewise for the type test.
1447 IRBuilder<> CallB((Preds.size() == 1 && !HasNonCallUses) ? Preds[0] : CI);
1448 CallInst *TypeTestCall = CallB.CreateCall(TypeTestFunc, {Ptr, TypeIdValue});
1450 for (Instruction *Pred : Preds) {
1451 Pred->replaceAllUsesWith(TypeTestCall);
1452 Pred->eraseFromParent();
1455 // We have already erased any extractvalue instructions that refer to the
1456 // intrinsic call, but the intrinsic may have other non-extractvalue uses
1457 // (although this is unlikely). In that case, explicitly build a pair and
1459 if (!CI->use_empty()) {
1460 Value *Pair = UndefValue::get(CI->getType());
1462 Pair = B.CreateInsertValue(Pair, LoadedValue, {0});
1463 Pair = B.CreateInsertValue(Pair, TypeTestCall, {1});
1464 CI->replaceAllUsesWith(Pair);
1467 // The number of unsafe uses is initially the number of uses.
1468 auto &NumUnsafeUses = NumUnsafeUsesForTypeTest[TypeTestCall];
1469 NumUnsafeUses = DevirtCalls.size();
1471 // If the function pointer has a non-call user, we cannot eliminate the type
1472 // check, as one of those users may eventually call the pointer. Increment
1473 // the unsafe use count to make sure it cannot reach zero.
1476 for (DevirtCallSite Call : DevirtCalls) {
1477 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS,
1481 CI->eraseFromParent();
1485 void DevirtModule::importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo) {
1486 const TypeIdSummary *TidSummary =
1487 ImportSummary->getTypeIdSummary(cast<MDString>(Slot.TypeID)->getString());
1490 auto ResI = TidSummary->WPDRes.find(Slot.ByteOffset);
1491 if (ResI == TidSummary->WPDRes.end())
1493 const WholeProgramDevirtResolution &Res = ResI->second;
1495 if (Res.TheKind == WholeProgramDevirtResolution::SingleImpl) {
1496 // The type of the function in the declaration is irrelevant because every
1497 // call site will cast it to the correct type.
1498 auto *SingleImpl = M.getOrInsertFunction(
1499 Res.SingleImplName, Type::getVoidTy(M.getContext()));
1501 // This is the import phase so we should not be exporting anything.
1502 bool IsExported = false;
1503 applySingleImplDevirt(SlotInfo, SingleImpl, IsExported);
1504 assert(!IsExported);
1507 for (auto &CSByConstantArg : SlotInfo.ConstCSInfo) {
1508 auto I = Res.ResByArg.find(CSByConstantArg.first);
1509 if (I == Res.ResByArg.end())
1511 auto &ResByArg = I->second;
1512 // FIXME: We should figure out what to do about the "function name" argument
1513 // to the apply* functions, as the function names are unavailable during the
1514 // importing phase. For now we just pass the empty string. This does not
1515 // impact correctness because the function names are just used for remarks.
1516 switch (ResByArg.TheKind) {
1517 case WholeProgramDevirtResolution::ByArg::UniformRetVal:
1518 applyUniformRetValOpt(CSByConstantArg.second, "", ResByArg.Info);
1520 case WholeProgramDevirtResolution::ByArg::UniqueRetVal: {
1521 Constant *UniqueMemberAddr =
1522 importGlobal(Slot, CSByConstantArg.first, "unique_member");
1523 applyUniqueRetValOpt(CSByConstantArg.second, "", ResByArg.Info,
1527 case WholeProgramDevirtResolution::ByArg::VirtualConstProp: {
1528 Constant *Byte = importConstant(Slot, CSByConstantArg.first, "byte",
1529 Int32Ty, ResByArg.Byte);
1530 Constant *Bit = importConstant(Slot, CSByConstantArg.first, "bit", Int8Ty,
1532 applyVirtualConstProp(CSByConstantArg.second, "", Byte, Bit);
1540 if (Res.TheKind == WholeProgramDevirtResolution::BranchFunnel) {
1541 auto *JT = M.getOrInsertFunction(getGlobalName(Slot, {}, "branch_funnel"),
1542 Type::getVoidTy(M.getContext()));
1543 bool IsExported = false;
1544 applyICallBranchFunnel(SlotInfo, JT, IsExported);
1545 assert(!IsExported);
1549 void DevirtModule::removeRedundantTypeTests() {
1550 auto True = ConstantInt::getTrue(M.getContext());
1551 for (auto &&U : NumUnsafeUsesForTypeTest) {
1552 if (U.second == 0) {
1553 U.first->replaceAllUsesWith(True);
1554 U.first->eraseFromParent();
1559 bool DevirtModule::run() {
1560 Function *TypeTestFunc =
1561 M.getFunction(Intrinsic::getName(Intrinsic::type_test));
1562 Function *TypeCheckedLoadFunc =
1563 M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load));
1564 Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume));
1566 // If only some of the modules were split, we cannot correctly handle
1567 // code that contains type tests or type checked loads.
1568 if ((ExportSummary && ExportSummary->partiallySplitLTOUnits()) ||
1569 (ImportSummary && ImportSummary->partiallySplitLTOUnits())) {
1570 if ((TypeTestFunc && !TypeTestFunc->use_empty()) ||
1571 (TypeCheckedLoadFunc && !TypeCheckedLoadFunc->use_empty()))
1572 report_fatal_error("inconsistent LTO Unit splitting with llvm.type.test "
1573 "or llvm.type.checked.load");
1577 // Normally if there are no users of the devirtualization intrinsics in the
1578 // module, this pass has nothing to do. But if we are exporting, we also need
1579 // to handle any users that appear only in the function summaries.
1580 if (!ExportSummary &&
1581 (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc ||
1582 AssumeFunc->use_empty()) &&
1583 (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty()))
1586 if (TypeTestFunc && AssumeFunc)
1587 scanTypeTestUsers(TypeTestFunc, AssumeFunc);
1589 if (TypeCheckedLoadFunc)
1590 scanTypeCheckedLoadUsers(TypeCheckedLoadFunc);
1592 if (ImportSummary) {
1593 for (auto &S : CallSlots)
1594 importResolution(S.first, S.second);
1596 removeRedundantTypeTests();
1598 // The rest of the code is only necessary when exporting or during regular
1599 // LTO, so we are done.
1603 // Rebuild type metadata into a map for easy lookup.
1604 std::vector<VTableBits> Bits;
1605 DenseMap<Metadata *, std::set<TypeMemberInfo>> TypeIdMap;
1606 buildTypeIdentifierMap(Bits, TypeIdMap);
1607 if (TypeIdMap.empty())
1610 // Collect information from summary about which calls to try to devirtualize.
1611 if (ExportSummary) {
1612 DenseMap<GlobalValue::GUID, TinyPtrVector<Metadata *>> MetadataByGUID;
1613 for (auto &P : TypeIdMap) {
1614 if (auto *TypeId = dyn_cast<MDString>(P.first))
1615 MetadataByGUID[GlobalValue::getGUID(TypeId->getString())].push_back(
1619 for (auto &P : *ExportSummary) {
1620 for (auto &S : P.second.SummaryList) {
1621 auto *FS = dyn_cast<FunctionSummary>(S.get());
1624 // FIXME: Only add live functions.
1625 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) {
1626 for (Metadata *MD : MetadataByGUID[VF.GUID]) {
1627 CallSlots[{MD, VF.Offset}]
1628 .CSInfo.markSummaryHasTypeTestAssumeUsers();
1631 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) {
1632 for (Metadata *MD : MetadataByGUID[VF.GUID]) {
1633 CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS);
1636 for (const FunctionSummary::ConstVCall &VC :
1637 FS->type_test_assume_const_vcalls()) {
1638 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) {
1639 CallSlots[{MD, VC.VFunc.Offset}]
1640 .ConstCSInfo[VC.Args]
1641 .markSummaryHasTypeTestAssumeUsers();
1644 for (const FunctionSummary::ConstVCall &VC :
1645 FS->type_checked_load_const_vcalls()) {
1646 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) {
1647 CallSlots[{MD, VC.VFunc.Offset}]
1648 .ConstCSInfo[VC.Args]
1649 .addSummaryTypeCheckedLoadUser(FS);
1656 // For each (type, offset) pair:
1657 bool DidVirtualConstProp = false;
1658 std::map<std::string, Function*> DevirtTargets;
1659 for (auto &S : CallSlots) {
1660 // Search each of the members of the type identifier for the virtual
1661 // function implementation at offset S.first.ByteOffset, and add to
1663 std::vector<VirtualCallTarget> TargetsForSlot;
1664 if (tryFindVirtualCallTargets(TargetsForSlot, TypeIdMap[S.first.TypeID],
1665 S.first.ByteOffset)) {
1666 WholeProgramDevirtResolution *Res = nullptr;
1667 if (ExportSummary && isa<MDString>(S.first.TypeID))
1668 Res = &ExportSummary
1669 ->getOrInsertTypeIdSummary(
1670 cast<MDString>(S.first.TypeID)->getString())
1671 .WPDRes[S.first.ByteOffset];
1673 if (!trySingleImplDevirt(TargetsForSlot, S.second, Res)) {
1674 DidVirtualConstProp |=
1675 tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first);
1677 tryICallBranchFunnel(TargetsForSlot, S.second, Res, S.first);
1680 // Collect functions devirtualized at least for one call site for stats.
1682 for (const auto &T : TargetsForSlot)
1684 DevirtTargets[T.Fn->getName()] = T.Fn;
1687 // CFI-specific: if we are exporting and any llvm.type.checked.load
1688 // intrinsics were *not* devirtualized, we need to add the resulting
1689 // llvm.type.test intrinsics to the function summaries so that the
1690 // LowerTypeTests pass will export them.
1691 if (ExportSummary && isa<MDString>(S.first.TypeID)) {
1693 GlobalValue::getGUID(cast<MDString>(S.first.TypeID)->getString());
1694 for (auto FS : S.second.CSInfo.SummaryTypeCheckedLoadUsers)
1695 FS->addTypeTest(GUID);
1696 for (auto &CCS : S.second.ConstCSInfo)
1697 for (auto FS : CCS.second.SummaryTypeCheckedLoadUsers)
1698 FS->addTypeTest(GUID);
1702 if (RemarksEnabled) {
1703 // Generate remarks for each devirtualized function.
1704 for (const auto &DT : DevirtTargets) {
1705 Function *F = DT.second;
1707 using namespace ore;
1708 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, "Devirtualized", F)
1710 << NV("FunctionName", F->getName()));
1714 removeRedundantTypeTests();
1716 // Rebuild each global we touched as part of virtual constant propagation to
1717 // include the before and after bytes.
1718 if (DidVirtualConstProp)
1719 for (VTableBits &B : Bits)