[llvm] Port Swift's merge function pass to llvm: merging functions that differ in constants (PR #68235)
Manman Ren via llvm-commits
llvm-commits at lists.llvm.org
Mon Oct 23 10:04:28 PDT 2023
================
@@ -0,0 +1,1430 @@
+//===--- MergeFunctionsIgnoringConst.cpp - Merge functions ----------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass looks for similar functions that are mergeable and folds them.
+// The implementation is similar to LLVM's MergeFunctions pass. Instead of
+// merging identical functions, it merges functions which only differ by a few
+// constants in certain instructions.
+// This is copied from Swift's implementation.
+// TODO: We should generalize this pass and share it with Swift's
+// implementation.
+//
+// This pass should run after LLVM's MergeFunctions pass, because it works best
+// if there are no _identical_ functions in the module.
+// Note: it would also work for identical functions but could produce more
+// code overhead than the LLVM pass.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Transforms/IPO/MergeFunctionsIgnoringConst.h"
+// #include "llvm/Transforms/Utils/GlobalMergeFunctions.h"
+#include "llvm/ADT/DenseSet.h"
+#include "llvm/ADT/FoldingSet.h"
+#include "llvm/ADT/Hashing.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Transforms/Utils/FunctionComparatorIgnoringConst.h"
+// #include "llvm/ADT/Triple.h"
+#include "llvm/Analysis/ObjCARCUtil.h"
+#include "llvm/ADT/StableHashing.h"
+#include "llvm/IR/Attributes.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/DebugInfoMetadata.h"
+// #include "llvm/IR/GlobalPtrAuthInfo.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/InlineAsm.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Operator.h"
+#include "llvm/IR/StructuralHash.h"
+#include "llvm/IR/ValueHandle.h"
+#include "llvm/IR/ValueMap.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/FileSystem.h"
+#include "llvm/Support/Regex.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Transforms/IPO.h"
+#include <vector>
+
+using namespace llvm;
+
+#define DEBUG_TYPE "mergefunc-ignoring-const"
+
+STATISTIC(NumFunctionsMergedIgnoringConst, "Number of functions merged");
+STATISTIC(NumThunksWrittenIgnoringConst, "Number of thunks generated");
+
+static cl::opt<bool>
+ EnableMergeFunc2("enable-merge-func2", cl::init(false), cl::Hidden,
+ cl::desc("Enable more aggressive function merger"));
+
+static cl::opt<unsigned> NumFunctionsIgnoringConstForSanityCheck(
+ "mergefunc-ignoringconst-sanity",
+ cl::desc("How many functions in module could be used for "
+ "MergeFunctionsIgnoringConst pass sanity check. "
+ "'0' disables this check. Works only with '-debug' key."),
+ cl::init(0), cl::Hidden);
+
+static cl::opt<unsigned> IgnoringConstMergeThreshold(
+ "mergefunc-ignoringconst-threshold",
+ cl::desc("Functions larger than the threshold are considered for merging."
+ "'0' disables function merging at all."),
+ cl::init(15), cl::Hidden);
+
+cl::opt<bool> UseLinkOnceODRLinkageMerging(
+ "use-linkonceodr-linkage-merging", cl::init(false), cl::Hidden,
+ cl::desc(
+ "Use LinkeOnceODR linkage to deduplicate the identical merged function "
+ "(default = off)"));
+
+cl::opt<bool> NoInlineForMergedFunction(
+ "no-inline-merged-function", cl::init(false), cl::Hidden,
+ cl::desc("set noinline for merged function (default = off)"));
+
+static cl::opt<bool>
+ CastArrayType("merge-cast-array-type", cl::init(false), cl::Hidden,
+ cl::desc("support for casting array type (default = off)"));
+
+static cl::opt<bool> IgnoreMusttailFunction(
+ "ignore-musttail-function", cl::init(false), cl::Hidden,
+ cl::desc(
+ "ignore functions containing callsites with musttail (default = off)"));
+
+static cl::opt<bool> AlwaysCallThunk(
+ "merge-always-call-thunk", cl::init(false), cl::Hidden,
+ cl::desc(
+ "do not replace callsites and always emit a thunk (default = off)"));
+
+static cl::list<std::string> MergeBlockRegexFilters(
+ "merge-block-regex", cl::Optional,
+ cl::desc("Block functions from merging if they match the given "
+ "regular expression"),
+ cl::ZeroOrMore);
+
+static cl::list<std::string> MergeAllowRegexFilters(
+ "merge-allow-regex", cl::Optional,
+ cl::desc("Allow functions from merging if they match the given "
+ "regular expression"),
+ cl::ZeroOrMore);
+
+bool isEligibleInstrunctionForConstantSharing(const Instruction *I) {
+ switch (I->getOpcode()) {
+ case Instruction::Load:
+ case Instruction::Store:
+ case Instruction::Call:
+ return true;
+ default: {
+ if (EnableMergeFunc2 && I->getOpcode() == Instruction::Invoke)
+ return true;
+ return false;
+ }
+ }
+}
+
+/// Returns true if the \opIdx operand of \p CI is the callee operand.
+static bool isCalleeOperand(const CallBase *CI, unsigned opIdx) {
+ return &CI->getCalledOperandUse() == &CI->getOperandUse(opIdx);
+}
+
+static bool canParameterizeCallOperand(const CallBase *CI, unsigned opIdx) {
+ if (CI->isInlineAsm())
+ return false;
+ Function *Callee = CI->getCalledOperand()
+ ? dyn_cast_or_null<Function>(
+ CI->getCalledOperand()->stripPointerCasts())
+ : nullptr;
+ if (Callee) {
+ if (Callee->isIntrinsic())
+ return false;
+ // objc_msgSend stubs must be called, and can't have their address taken.
+ if (Callee->getName().startswith("objc_msgSend$"))
+ return false;
+ }
+ if (isCalleeOperand(CI, opIdx) &&
+ CI->getOperandBundle(LLVMContext::OB_ptrauth).has_value()) {
+ // The operand is the callee and it has already been signed. Ignore this
+ // because we cannot add another ptrauth bundle to the call instruction.
+ return false;
+ }
+ return true;
+}
+
+bool isEligibleOperandForConstantSharing(const Instruction *I, unsigned OpIdx) {
+ assert(OpIdx < I->getNumOperands() && "Invalid operand index");
+
+ if (!isEligibleInstrunctionForConstantSharing(I))
+ return false;
+
+ auto Opnd = I->getOperand(OpIdx);
+ if (!isa<Constant>(Opnd))
+ return false;
+
+ if (const auto *CI = dyn_cast<CallBase>(I))
+ return canParameterizeCallOperand(CI, OpIdx);
+
+ return true;
+}
+
+namespace {
+
+/// MergeFuncIgnoringConst finds functions which only differ by constants in
+/// certain instructions, e.g. resulting from specialized functions of layout
+/// compatible types.
+/// Such functions are merged by replacing the differing constants by a
+/// parameter. The original functions are replaced by thunks which call the
+/// merged function with the specific argument constants.
+///
+class MergeFuncIgnoringConstImpl { // : public ModulePass {
+public:
+ MergeFuncIgnoringConstImpl() : FnTree(FunctionNodeCmp(&GlobalNumbers)) {}
+
+ MergeFuncIgnoringConstImpl(bool ptrAuthEnabled, unsigned ptrAuthKey)
+ : FnTree(FunctionNodeCmp(&GlobalNumbers)), ptrAuthOptionsSet(true),
+ ptrAuthEnabled(ptrAuthEnabled), ptrAuthKey(ptrAuthKey) {}
+
+ bool runImpl(Module &M);
+
+private:
+ struct FunctionEntry;
+
+ /// Describes the set of functions which are considered as "equivalent" (i.e.
+ /// only differing by some constants).
+ struct EquivalenceClass {
+ /// The single-linked list of all functions which are a member of this
+ /// equivalence class.
+ FunctionEntry *First;
+
+ /// A very cheap hash, used to early exit if functions do not match.
+ llvm::IRHash Hash;
+
+ public:
+ // Note the hash is recalculated potentially multiple times, but it is
+ // cheap.
+ EquivalenceClass(FunctionEntry *First)
+ : First(First), Hash(StructuralHash(*First->F)) {
+ assert(!First->Next);
+ }
+ };
+
+ /// The function comparison operator is provided here so that FunctionNodes do
+ /// not need to become larger with another pointer.
+ class FunctionNodeCmp {
+ GlobalNumberState *GlobalNumbers;
+
+ public:
+ FunctionNodeCmp(GlobalNumberState *GN) : GlobalNumbers(GN) {}
+ bool operator()(const EquivalenceClass &LHS,
+ const EquivalenceClass &RHS) const {
+ // Order first by hashes, then full function comparison.
+ if (LHS.Hash != RHS.Hash)
+ return LHS.Hash < RHS.Hash;
+ FunctionComparatorIgnoringConst FCmp(LHS.First->F, RHS.First->F,
+ GlobalNumbers);
+ return FCmp.compareIgnoringConsts() == -1;
+ }
+ };
+ using FnTreeType = std::set<EquivalenceClass, FunctionNodeCmp>;
+
+ ///
+ struct FunctionEntry {
+ FunctionEntry(Function *F, FnTreeType::iterator I)
+ : F(F), Next(nullptr), numUnhandledCallees(0), TreeIter(I),
+ isMerged(false) {}
+
+ /// Back-link to the function.
+ AssertingVH<Function> F;
+
+ /// The next function in its equivalence class.
+ FunctionEntry *Next;
+
+ /// The number of not-yet merged callees. Used to process the merging in
+ /// bottom-up call order.
+ /// This is only valid in the first entry of an equivalence class. The
+ /// counts of all functions in an equivalence class are accumulated in the
+ /// first entry.
+ int numUnhandledCallees;
+
+ /// The iterator of the function's equivalence class in the FnTree.
+ /// It's FnTree.end() if the function is not in an equivalence class.
+ FnTreeType::iterator TreeIter;
+
+ /// True if this function is already a thunk, calling the merged function.
+ bool isMerged;
+ };
+
+ /// Describes an operator of a specific instruction.
+ struct OpLocation {
+ Instruction *I;
+ unsigned OpIndex;
+ };
+
+ /// Information for a function. Used during merging.
+ struct FunctionInfo {
+
+ FunctionInfo(Function *F)
+ : F(F), CurrentInst(nullptr), NumParamsNeeded(0) {}
+
+ void init() {
+ CurrentInst = &*F->begin()->begin();
+ NumParamsNeeded = 0;
+ }
+
+ /// Advances the current instruction to the next instruction.
+ void nextInst() {
+ assert(CurrentInst);
+ if (CurrentInst->isTerminator()) {
+ auto BlockIter = std::next(CurrentInst->getParent()->getIterator());
+ if (BlockIter == F->end()) {
+ CurrentInst = nullptr;
+ return;
+ }
+ CurrentInst = &*BlockIter->begin();
+ return;
+ }
+ CurrentInst = &*std::next(CurrentInst->getIterator());
+ }
+
+ /// Returns true if the operand \p OpIdx of the current instruction is the
+ /// callee of a call, which needs to be signed if passed as a parameter.
+ bool needsPointerSigning(unsigned OpIdx) const {
+ if (auto *CI = dyn_cast<CallInst>(CurrentInst))
+ return isCalleeOperand(CI, OpIdx);
+ return false;
+ }
+
+ Function *F;
+
+ /// The current instruction while iterating over all instructions.
+ Instruction *CurrentInst;
+
+ /// Roughly the number of parameters needed if this function would be
+ /// merged with the first function of the equivalence class.
+ int NumParamsNeeded;
+ };
+
+ using FunctionInfos = SmallVector<FunctionInfo, 8>;
+
+ /// Describes a parameter which we create to parameterize the merged function.
+ struct ParamInfo {
+ /// The value of the parameter for all the functions in the equivalence
+ /// class.
+ SmallVector<Constant *, 8> Values;
+
+ /// All uses of the parameter in the merged function.
+ SmallVector<OpLocation, 16> Uses;
+
+ /// The discriminator for pointer signing.
+ /// Only not null if needsPointerSigning is true.
+ ConstantInt *discriminator = nullptr;
+
+ /// True if the value is a callee function, which needs to be signed if
+ /// passed as a parameter.
+ bool needsPointerSigning = false;
+
+ /// Checks if this parameter can be used to describe an operand in all
+ /// functions of the equivalence class. Returns true if all values match
+ /// the specific instruction operands in all functions.
+ bool matches(const FunctionInfos &FInfos, unsigned OpIdx,
+ bool ptrAuthEnabled) const {
+ unsigned NumFuncs = FInfos.size();
+ assert(Values.size() == NumFuncs);
+ if (ptrAuthEnabled &&
+ needsPointerSigning != FInfos[0].needsPointerSigning(OpIdx)) {
+ return false;
+ }
+ for (unsigned Idx = 0; Idx < NumFuncs; ++Idx) {
+ const FunctionInfo &FI = FInfos[Idx];
+ Constant *C = cast<Constant>(FI.CurrentInst->getOperand(OpIdx));
+ if (Values[Idx] != C)
+ return false;
+ }
+ return true;
+ }
+
+ /// Computes the discriminator for pointer signing.
+ void computeDiscriminator(LLVMContext &Context) {
+ assert(needsPointerSigning);
+ assert(!discriminator);
+
+ /// Get a hash from the concatenated function names.
+ /// The hash is deterministic, because the order of values depends on the
+ /// order of functions in the module, which is itself deterministic.
+ /// Note that the hash is not part of the ABI, because it's purly used
+ /// for pointer authentication between a module-private caller-callee
+ /// pair.
+ std::string concatenatedCalleeNames;
+ for (Constant *value : Values) {
+ if (auto *GO = dyn_cast<GlobalObject>(value))
+ concatenatedCalleeNames += GO->getName();
+ }
+ uint64_t rawHash = stable_hash_combine_string(concatenatedCalleeNames);
+ IntegerType *discrTy = Type::getInt64Ty(Context);
+ discriminator = ConstantInt::get(discrTy, (rawHash % 0xFFFF) + 1);
+ }
+ };
+
+ using ParamInfos = SmallVector<ParamInfo, 16>;
+
+ Module *module = nullptr;
+ ModuleSummaryIndex *ExportSummary;
+ const ModuleSummaryIndex *ImportSummary;
+
+ GlobalNumberState GlobalNumbers;
+
+ /// A work queue of functions that may have been modified and should be
+ /// analyzed again.
+ std::vector<WeakTrackingVH> Deferred;
+
+ /// The set of all distinct functions. Use the insert() and remove() methods
+ /// to modify it. The map allows efficient lookup and deferring of Functions.
+ FnTreeType FnTree;
+
+ ValueMap<Function *, FunctionEntry *> FuncEntries;
+
+ // Maps a function-pointer / discriminator pair to a corresponding global in
+ // the llvm.ptrauth section.
+ // This map is used as a cache to not create ptrauth globals twice.
+ DenseMap<std::pair<Constant *, ConstantInt *>, Constant *> ptrAuthGlobals;
+
+ /// If true, ptrAuthEnabled and ptrAuthKey are valid.
+ bool ptrAuthOptionsSet = false;
+
+ /// True if the architecture has pointer authentication enabled.
+ bool ptrAuthEnabled = false;
+
+ /// The key for pointer authentication.
+ unsigned ptrAuthKey = 0;
+
+ FunctionEntry *getEntry(Function *F) const { return FuncEntries.lookup(F); }
+
+ bool isInEquivalenceClass(FunctionEntry *FE) const {
+ if (FE->TreeIter != FnTree.end()) {
+ return true;
+ }
+ assert(!FE->Next);
+ assert(FE->numUnhandledCallees == 0);
+ return false;
+ }
+
+ /// Checks the rules of order relation introduced among functions set.
+ /// Returns true, if sanity check has been passed, and false if failed.
+ bool doSanityCheck(std::vector<WeakTrackingVH> &Worklist);
+
+ /// Updates the numUnhandledCallees of all user functions of the equivalence
+ /// class containing \p FE by \p Delta.
+ void updateUnhandledCalleeCount(FunctionEntry *FE, int Delta);
+
+ bool tryMergeEquivalenceClass(FunctionEntry *FirstInClass);
+
+ FunctionInfo removeFuncWithMostParams(FunctionInfos &FInfos);
+
+ bool deriveParams(ParamInfos &Params, FunctionInfos &FInfos,
+ unsigned maxParams);
+
+ bool numOperandsDiffer(FunctionInfos &FInfos);
+
+ bool constsDiffer(const FunctionInfos &FInfos, unsigned OpIdx);
+
+ bool tryMapToParameter(FunctionInfos &FInfos, unsigned OpIdx,
+ ParamInfos &Params, unsigned maxParams);
+
+ void replaceCallWithAddedPtrAuth(CallInst *origCall, Value *newCallee,
+ ConstantInt *discriminator);
+
+ void mergeWithParams(const FunctionInfos &FInfos, ParamInfos &Params);
+ static void dumpMergeInfo(const FunctionInfos &FInfos, unsigned);
+
+ void removeEquivalenceClassFromTree(FunctionEntry *FE);
+
+ void writeThunk(Function *ToFunc, Function *Thunk, const ParamInfos &Params,
+ unsigned FuncIdx);
+
+ bool isPtrAuthEnabled() const {
+ // TODO: fix pointer authentication
+ // assert(ptrAuthOptionsSet);
+ return ptrAuthEnabled;
+ }
+
+ ConstantInt *getPtrAuthKey() {
+ // TODO: fix pointer authentication
+ // assert(isPtrAuthEnabled());
+ return ConstantInt::get(Type::getInt32Ty(module->getContext()), ptrAuthKey);
+ }
+
+ /// Returns the value of function \p FuncIdx, and signes it if required.
+ Constant *getSignedValue(const ParamInfo &PI, unsigned FuncIdx) {
+ Constant *value = PI.Values[FuncIdx];
+ if (!PI.needsPointerSigning)
+ return value;
+
+ auto lookupKey = std::make_pair(value, PI.discriminator);
+ Constant *&ptrAuthGlobal = ptrAuthGlobals[lookupKey];
+ if (!ptrAuthGlobal) {
+#if 0
+ ptrAuthGlobal = GlobalPtrAuthInfo::create(
+ *module, value, getPtrAuthKey(),
+ ConstantInt::get(PI.discriminator->getType(), 0), PI.discriminator);
+#endif
+ }
+ return ptrAuthGlobal;
+ }
+
+ /// Replace all direct calls of Old with calls of New. Will bitcast New if
+ /// necessary to make types match.
+ bool replaceDirectCallers(Function *Old, Function *New,
+ const ParamInfos &Params, unsigned FuncIdx);
+};
+
+#if 0
+class MergeFuncIgnoringConst : public ModulePass {
+public:
+ static char ID;
+ /// True if the architecture has pointer authentication enabled.
+ bool ptrAuthEnabled = false;
+
+ /// The key for pointer authentication.
+ unsigned ptrAuthKey = 0;
+ ModuleSummaryIndex *ExportSummary;
+ const ModuleSummaryIndex *ImportSummary;
+
+ MergeFuncIgnoringConst() : ModulePass(ID) {
+ initializeMergeFuncIgnoringConstPass(*llvm::PassRegistry::getPassRegistry());
+ }
+ MergeFuncIgnoringConst(bool ptrAuthEnabled, unsigned ptrAuthKey)
+ : ModulePass(ID), ptrAuthEnabled(ptrAuthEnabled), ptrAuthKey(ptrAuthKey) {
+ initializeMergeFuncIgnoringConstPass(*llvm::PassRegistry::getPassRegistry());
+ }
+ bool runOnModule(Module &M) override;
+};
+#endif
+
+} // end anonymous namespace
+
+#if 0
+char MergeFuncIgnoringConst::ID = 0;
+INITIALIZE_PASS_BEGIN(MergeFuncIgnoringConst, "merge-func-ignoring-const",
+ "merge function pass ignoring const", false, false)
+INITIALIZE_PASS_END(MergeFuncIgnoringConst, "merge-func-ignoring-const",
+ "merge function pass ignoring const", false, false)
+#endif
+bool MergeFuncIgnoringConstImpl::doSanityCheck(
+ std::vector<WeakTrackingVH> &Worklist) {
+ if (const unsigned Max = NumFunctionsIgnoringConstForSanityCheck) {
+ unsigned TripleNumber = 0;
+ bool Valid = true;
+
+ dbgs() << "MERGEFUNC-SANITY: Started for first " << Max << " functions.\n";
+
+ unsigned i = 0;
+ for (std::vector<WeakTrackingVH>::iterator I = Worklist.begin(),
+ E = Worklist.end();
+ I != E && i < Max; ++I, ++i) {
+ unsigned j = i;
+ for (std::vector<WeakTrackingVH>::iterator J = I; J != E && j < Max;
+ ++J, ++j) {
+ Function *F1 = cast<Function>(*I);
+ Function *F2 = cast<Function>(*J);
+ int Res1 = FunctionComparatorIgnoringConst(F1, F2, &GlobalNumbers)
+ .compareIgnoringConsts();
+ int Res2 = FunctionComparatorIgnoringConst(F2, F1, &GlobalNumbers)
+ .compareIgnoringConsts();
+
+ // If F1 <= F2, then F2 >= F1, otherwise report failure.
+ if (Res1 != -Res2) {
+ dbgs() << "MERGEFUNC-SANITY: Non-symmetric; triple: " << TripleNumber
+ << "\n";
+ LLVM_DEBUG(F1->dump());
+ LLVM_DEBUG(F2->dump());
+ Valid = false;
+ }
+
+ if (Res1 == 0)
+ continue;
+
+ unsigned k = j;
+ for (std::vector<WeakTrackingVH>::iterator K = J; K != E && k < Max;
+ ++k, ++K, ++TripleNumber) {
+ if (K == J)
+ continue;
+
+ Function *F3 = cast<Function>(*K);
+ int Res3 = FunctionComparatorIgnoringConst(F1, F3, &GlobalNumbers)
+ .compareIgnoringConsts();
+ int Res4 = FunctionComparatorIgnoringConst(F2, F3, &GlobalNumbers)
+ .compareIgnoringConsts();
+
+ bool Transitive = true;
+
+ if (Res1 != 0 && Res1 == Res4) {
+ // F1 > F2, F2 > F3 => F1 > F3
+ Transitive = Res3 == Res1;
+ } else if (Res3 != 0 && Res3 == -Res4) {
+ // F1 > F3, F3 > F2 => F1 > F2
+ Transitive = Res3 == Res1;
+ } else if (Res4 != 0 && -Res3 == Res4) {
+ // F2 > F3, F3 > F1 => F2 > F1
+ Transitive = Res4 == -Res1;
+ }
+
+ if (!Transitive) {
+ dbgs() << "MERGEFUNC-SANITY: Non-transitive; triple: "
+ << TripleNumber << "\n";
+ dbgs() << "Res1, Res3, Res4: " << Res1 << ", " << Res3 << ", "
+ << Res4 << "\n";
+ LLVM_DEBUG(F1->dump());
+ LLVM_DEBUG(F2->dump());
+ LLVM_DEBUG(F3->dump());
+ Valid = false;
+ }
+ }
+ }
+ }
+
+ dbgs() << "MERGEFUNC-SANITY: " << (Valid ? "Passed." : "Failed.") << "\n";
+ return Valid;
+ }
+ return true;
+}
+
+/// Returns true if functions containing calls to \p F may be merged together.
+static bool mayMergeCallsToFunction(Function &F) {
+ StringRef Name = F.getName();
+
+ // Calls to dtrace probes must generate unique patchpoints.
+ if (Name.startswith("__dtrace"))
+ return false;
+
+ return true;
+}
+
+/// Returns the benefit, which is approximately the size of the function.
+/// Return 0, if the function should not be merged.
+static unsigned getBenefit(Function *F) {
+ unsigned Benefit = 0;
+
+ // We don't want to merge very small functions, because the overhead of
+ // adding creating thunks and/or adding parameters to the call sites
+ // outweighs the benefit.
+ for (BasicBlock &BB : *F) {
+ for (Instruction &I : BB) {
+ if (CallBase *CB = dyn_cast<CallBase>(&I)) {
+ Function *Callee = CB->getCalledFunction();
+ if (Callee && !mayMergeCallsToFunction(*Callee))
+ return 0;
+ if (!Callee || !Callee->isIntrinsic()) {
+ Benefit += 5;
+ continue;
+ }
+ }
+ Benefit += 1;
+ }
+ }
+ return Benefit;
+}
+
+/// Returns true if function \p F is eligible for merging.
+bool isEligibleFunction(Function *F) {
+ if (F->isDeclaration())
+ return false;
+
+ if (F->hasFnAttribute(llvm::Attribute::NoMerge))
+ return false;
+
+ if (F->hasAvailableExternallyLinkage()) {
+ return false;
+ }
+
+ if (F->getFunctionType()->isVarArg()) {
+ return false;
+ }
+
+ // Check against blocklist.
+ if (!MergeBlockRegexFilters.empty()) {
+ StringRef FuncName = F->getName();
+ for (const auto &tRegex : MergeBlockRegexFilters)
+ if (Regex(tRegex).match(FuncName)) {
+ return false;
+ }
+ }
+ // Check against allowlist
+ if (!MergeAllowRegexFilters.empty()) {
+ StringRef FuncName = F->getName();
+ bool found = false;
+ for (const auto &tRegex : MergeAllowRegexFilters)
+ if (Regex(tRegex).match(FuncName)) {
+ found = true;
+ break;
+ }
+ if (!found)
+ return false;
+ }
+
+ if (F->getCallingConv() == CallingConv::SwiftTail)
+ return false;
+
+ // if function contains callsites with musttail, if we merge
+ // it, the merged function will have the musttail callsite, but
+ // the number of parameters can change, thus the parameter count
+ // of the callsite will mismatch with the function itself.
+ if (IgnoreMusttailFunction) {
+ for (const BasicBlock &BB : *F) {
+ for (const Instruction &I : BB) {
+ const auto *CB = dyn_cast<CallBase>(&I);
+ if (CB && CB->isMustTailCall())
+ return false;
+ }
+ }
+ }
+
+ unsigned Benefit = getBenefit(F);
+ if (Benefit < IgnoringConstMergeThreshold) {
+ return false;
+ }
+
+ return true;
+}
+
+static bool runInternal(Module &M) {
+ return MergeFuncIgnoringConstImpl().runImpl(M);
+}
+
+// bool MergeFuncIgnoringConst::runOnModule(Module &M) { return runInternal(M);
+// }
+
+bool MergeFuncIgnoringConstImpl::runImpl(Module &M) {
+ if (IgnoringConstMergeThreshold == 0)
+ return false;
+
+ module = &M;
+
+#if 0
+ // TODO: fix pointer authentication
+ if (!ptrAuthOptionsSet) {
+ // If invoked from IRGen in the compiler, those options are already set.
+ // If invoked from swift-llvm-opt, derive the options from the target triple.
+ Triple triple(M.getTargetTriple());
+ ptrAuthEnabled = (triple.getSubArch() == Triple::AArch64SubArch_arm64e);
+ ptrAuthKey = (unsigned)clang::PointerAuthSchema::ARM8_3Key::ASIA;
+ ptrAuthOptionsSet = true;
+ }
+#endif
+
+ bool Changed = false;
+
+ // All functions in the module, ordered by hash. Functions with a unique
+ // hash value are easily eliminated.
+ std::vector<std::pair<llvm::IRHash, Function *>> HashedFuncs;
+
+ for (Function &Func : M) {
+ if (isEligibleFunction(&Func)) {
+ HashedFuncs.push_back({StructuralHash(Func), &Func});
+ }
+ }
+
+ std::stable_sort(HashedFuncs.begin(), HashedFuncs.end(),
+ [](const std::pair<llvm::IRHash, Function *> &a,
+ const std::pair<llvm::IRHash, Function *> &b) {
+ return a.first < b.first;
+ });
+
+ std::vector<FunctionEntry> FuncEntryStorage;
+ FuncEntryStorage.reserve(HashedFuncs.size());
+
+ auto S = HashedFuncs.begin();
+ for (auto I = HashedFuncs.begin(), IE = HashedFuncs.end(); I != IE; ++I) {
+
+ Function *F = I->second;
+ FuncEntryStorage.push_back(FunctionEntry(F, FnTree.end()));
+ FunctionEntry &FE = FuncEntryStorage.back();
+ FuncEntries[F] = &FE;
+
+ // If the hash value matches the previous value or the next one, we must
+ // consider merging it. Otherwise it is dropped and never considered again.
+ if ((I != S && std::prev(I)->first == I->first) ||
+ (std::next(I) != IE && std::next(I)->first == I->first)) {
+ Deferred.push_back(WeakTrackingVH(F));
+ }
+ }
+
+ do {
+ std::vector<WeakTrackingVH> Worklist;
+ Deferred.swap(Worklist);
+
+ LLVM_DEBUG(dbgs() << "======\nbuild tree: worklist-size=" << Worklist.size()
+ << '\n');
+ LLVM_DEBUG(doSanityCheck(Worklist));
+
+ SmallVector<FunctionEntry *, 8> FuncsToMerge;
+
+ // Insert all candidates into the Worklist.
+ for (WeakTrackingVH &I : Worklist) {
+ if (!I)
+ continue;
+ Function *F = cast<Function>(I);
+ FunctionEntry *FE = getEntry(F);
+ assert(!isInEquivalenceClass(FE));
+
+ std::pair<FnTreeType::iterator, bool> Result = FnTree.insert(FE);
+
+ FE->TreeIter = Result.first;
+ const EquivalenceClass &Eq = *Result.first;
+
+ if (Result.second) {
+ assert(Eq.First == FE);
+ LLVM_DEBUG(dbgs() << " new in tree: " << F->getName() << '\n');
+ } else {
+ assert(Eq.First != FE);
+ LLVM_DEBUG(dbgs() << " add to existing: " << F->getName() << '\n');
+ // Add the function to the existing equivalence class.
+ FE->Next = Eq.First->Next;
+ Eq.First->Next = FE;
+ // Schedule for merging if the function's equivalence class reaches the
+ // size of 2.
+ if (!FE->Next)
+ FuncsToMerge.push_back(Eq.First);
+ }
+ }
+ LLVM_DEBUG(dbgs() << "merge functions: tree-size=" << FnTree.size()
+ << '\n');
+
+ // Figure out the leaf functions. We want to do the merging in bottom-up
+ // call order. This ensures that we don't parameterize on callee function
+ // names if we don't have to (because the callee may be merged).
+ // Note that "leaf functions" refer to the sub-call-graph of functions which
+ // are in the FnTree.
+ for (FunctionEntry *ToMerge : FuncsToMerge) {
+ assert(isInEquivalenceClass(ToMerge));
+ updateUnhandledCalleeCount(ToMerge, 1);
+ }
+
+ // Check if there are any leaf functions at all.
+ bool LeafFound = false;
+ for (FunctionEntry *ToMerge : FuncsToMerge) {
+ if (ToMerge->numUnhandledCallees == 0)
+ LeafFound = true;
+ }
+ for (FunctionEntry *ToMerge : FuncsToMerge) {
+ if (isInEquivalenceClass(ToMerge)) {
+ // Only merge leaf functions (or all functions if all functions are in
+ // a call cycle).
+ if (ToMerge->numUnhandledCallees == 0 || !LeafFound) {
+ updateUnhandledCalleeCount(ToMerge, -1);
+ Changed |= tryMergeEquivalenceClass(ToMerge);
+ } else {
+ // Non-leaf functions (i.e. functions in a call cycle) may become
+ // leaf functions in the next iteration.
+ removeEquivalenceClassFromTree(ToMerge);
+ }
+ }
+ }
+ } while (!Deferred.empty());
+
+ FnTree.clear();
+ GlobalNumbers.clear();
+ FuncEntries.clear();
+ ptrAuthGlobals.clear();
+
+ return Changed;
+}
+
+void MergeFuncIgnoringConstImpl::updateUnhandledCalleeCount(FunctionEntry *FE,
+ int Delta) {
+ // Iterate over all functions of FE's equivalence class.
+ do {
+ for (Use &U : FE->F->uses()) {
+ if (auto *I = dyn_cast<Instruction>(U.getUser())) {
+ FunctionEntry *CallerFE = getEntry(I->getFunction());
+ if (CallerFE && CallerFE->TreeIter != FnTree.end()) {
+ // Accumulate the count in the first entry of the equivalence class.
+ FunctionEntry *Head = CallerFE->TreeIter->First;
+ Head->numUnhandledCallees += Delta;
+ }
+ }
+ }
+ FE = FE->Next;
+ } while (FE);
+}
+
+bool MergeFuncIgnoringConstImpl::tryMergeEquivalenceClass(
+ FunctionEntry *FirstInClass) {
+ // Build the FInfos vector from all functions in the equivalence class.
+ FunctionInfos FInfos;
+ FunctionEntry *FE = FirstInClass;
+ do {
+ FInfos.push_back(FunctionInfo(FE->F));
+ FE->isMerged = true;
+ FE = FE->Next;
+ } while (FE);
+ assert(FInfos.size() >= 2);
+
+ // Merged or not: in any case we remove the equivalence class from the FnTree.
+ removeEquivalenceClassFromTree(FirstInClass);
+
+ // Contains functions which differ too much from the first function (i.e.
+ // would need too many parameters).
+ FunctionInfos Removed;
+
+ bool Changed = false;
+ int Try = 0;
+
+ unsigned Benefit = getBenefit(FirstInClass->F);
+
+ // The bigger the function, the more parameters are allowed.
+ unsigned maxParams = std::max(4u, Benefit / 100);
+
+ // We need multiple tries if there are some functions in FInfos which differ
+ // too much from the first function in FInfos. But we limit the number of
+ // tries to a small number, because this is quadratic.
+ while (FInfos.size() >= 2 && Try++ < 4) {
+ ParamInfos Params;
+ bool Merged = deriveParams(Params, FInfos, maxParams);
+ if (Merged) {
+ mergeWithParams(FInfos, Params);
+ Changed = true;
+ } else {
+ // We ran out of parameters. Remove the function from the set which
+ // differs most from the first function.
+ Removed.push_back(removeFuncWithMostParams(FInfos));
+ }
+ if (Merged || FInfos.size() < 2) {
+ // Try again with the functions which were removed from the original set.
+ FInfos.swap(Removed);
+ Removed.clear();
+ }
+ }
+ return Changed;
+}
+
+/// Remove the function from \p FInfos which needs the most parameters. Add the
+/// removed function to
+MergeFuncIgnoringConstImpl::FunctionInfo
+MergeFuncIgnoringConstImpl::removeFuncWithMostParams(FunctionInfos &FInfos) {
+ FunctionInfos::iterator MaxIter = FInfos.end();
+ for (auto Iter = FInfos.begin(), End = FInfos.end(); Iter != End; ++Iter) {
+ if (MaxIter == FInfos.end() ||
+ Iter->NumParamsNeeded > MaxIter->NumParamsNeeded) {
+ MaxIter = Iter;
+ }
+ }
+ FunctionInfo Removed = *MaxIter;
+ FInfos.erase(MaxIter);
+ return Removed;
+}
+
+/// Finds the set of parameters which are required to merge the functions in
+/// \p FInfos.
+/// Returns true on success, i.e. the functions in \p FInfos can be merged with
+/// the parameters returned in \p Params.
+bool MergeFuncIgnoringConstImpl::deriveParams(ParamInfos &Params,
+ FunctionInfos &FInfos,
+ unsigned maxParams) {
+ for (FunctionInfo &FI : FInfos)
+ FI.init();
+
+ FunctionInfo &FirstFI = FInfos.front();
+
+ // Iterate over all instructions synchronously in all functions.
+ do {
+ if (isEligibleInstrunctionForConstantSharing(FirstFI.CurrentInst)) {
+
+ // Here we handle a rare corner case which needs to be explained:
+ // Usually the number of operands match, because otherwise the functions
+ // in FInfos would not be in the same equivalence class. There is only one
+ // exception to that: If the current instruction is a call to a function,
+ // which was merged in the previous iteration (in
+ // tryMergeEquivalenceClass) then the call could be replaced and has more
+ // arguments than the original call.
+ if (numOperandsDiffer(FInfos)) {
+ assert(isa<CallInst>(FirstFI.CurrentInst) &&
+ "only calls are expected to differ in number of operands");
+ return false;
+ }
+
+ for (unsigned OpIdx = 0, NumOps = FirstFI.CurrentInst->getNumOperands();
+ OpIdx != NumOps; ++OpIdx) {
+
+ if (constsDiffer(FInfos, OpIdx)) {
+ // This instruction has operands which differ in at least some
+ // functions. So we need to parameterize it.
+ if (!tryMapToParameter(FInfos, OpIdx, Params, maxParams)) {
+ // We ran out of parameters.
+ return false;
+ }
+ }
+ }
+ }
+ // Go to the next instruction in all functions.
+ for (FunctionInfo &FI : FInfos)
+ FI.nextInst();
+ } while (FirstFI.CurrentInst);
+
+ return true;
+}
+
+/// Returns true if the number of operands of the current instruction differs.
+bool MergeFuncIgnoringConstImpl::numOperandsDiffer(FunctionInfos &FInfos) {
+ unsigned numOps = FInfos[0].CurrentInst->getNumOperands();
+ for (const FunctionInfo &FI : ArrayRef<FunctionInfo>(FInfos).drop_front(1)) {
+ if (FI.CurrentInst->getNumOperands() != numOps)
+ return true;
+ }
+ return false;
+}
+
+/// Returns true if the \p OpIdx's constant operand in the current instruction
+/// does differ in any of the functions in \p FInfos.
+bool MergeFuncIgnoringConstImpl::constsDiffer(const FunctionInfos &FInfos,
+ unsigned OpIdx) {
+ Constant *CommonConst = nullptr;
+
+ for (const FunctionInfo &FI : FInfos) {
+ Value *Op = FI.CurrentInst->getOperand(OpIdx);
+ if (auto *C = dyn_cast<Constant>(Op)) {
+ if (!CommonConst) {
+ CommonConst = C;
+ } else if (EnableMergeFunc2 && isa<ConstantPointerNull>(CommonConst) &&
+ isa<ConstantPointerNull>(C)) {
+ // if both are null pointer, and if they are different constants
+ // due to type, still treat them as the same.
+ } else if (C != CommonConst) {
+ return true;
+ }
+ }
+ }
+ return false;
+}
+
+/// Create a new parameter for differing operands or try to reuse an existing
+/// parameter.
+/// Returns true if a parameter could be created or found without exceeding the
+/// maximum number of parameters.
+bool MergeFuncIgnoringConstImpl::tryMapToParameter(FunctionInfos &FInfos,
+ unsigned OpIdx,
+ ParamInfos &Params,
+ unsigned maxParams) {
+ ParamInfo *Matching = nullptr;
+ // Try to find an existing parameter which exactly matches the differing
+ // operands of the current instruction.
+ for (ParamInfo &PI : Params) {
+ if (PI.matches(FInfos, OpIdx, isPtrAuthEnabled())) {
+ Matching = &PI;
+ break;
+ }
+ }
+ if (!Matching) {
+ // We need a new parameter.
+ // Check if we are within the limit.
+ if (Params.size() >= maxParams)
+ return false;
+
+ Params.resize(Params.size() + 1);
+ Matching = &Params.back();
+ // Store the constant values into the new parameter.
+ Constant *FirstC = cast<Constant>(FInfos[0].CurrentInst->getOperand(OpIdx));
+ for (FunctionInfo &FI : FInfos) {
+ Constant *C = cast<Constant>(FI.CurrentInst->getOperand(OpIdx));
+ Matching->Values.push_back(C);
+ if (C != FirstC)
+ FI.NumParamsNeeded += 1;
+ }
+ if (isPtrAuthEnabled())
+ Matching->needsPointerSigning = FInfos[0].needsPointerSigning(OpIdx);
+ }
+ /// Remember where the parameter is needed when we build our merged function.
+ Matching->Uses.push_back({FInfos[0].CurrentInst, OpIdx});
+ return true;
+}
+
+/// Copy \p origCall with a \p newCalle and add a ptrauth bundle with \p
+/// discriminator.
+void MergeFuncIgnoringConstImpl::replaceCallWithAddedPtrAuth(
+ CallInst *origCall, Value *newCallee, ConstantInt *discriminator) {
+ SmallVector<llvm::OperandBundleDef, 4> bundles;
+ origCall->getOperandBundlesAsDefs(bundles);
+ ConstantInt *key = getPtrAuthKey();
+ llvm::Value *bundleArgs[] = {key, discriminator};
+ bundles.emplace_back("ptrauth", bundleArgs);
+
+ SmallVector<llvm::Value *, 4> copiedArgs;
+ for (Value *op : origCall->args()) {
+ copiedArgs.push_back(op);
+ }
+
+ auto *newCall =
+ CallInst::Create(origCall->getFunctionType(), newCallee, copiedArgs,
+ bundles, origCall->getName(), origCall);
+ newCall->setAttributes(origCall->getAttributes());
+ newCall->setTailCallKind(origCall->getTailCallKind());
+ newCall->setCallingConv(origCall->getCallingConv());
+ origCall->replaceAllUsesWith(newCall);
+ origCall->eraseFromParent();
+}
+
+void MergeFuncIgnoringConstImpl::dumpMergeInfo(const FunctionInfos &FInfos,
+ unsigned paramSize) {
+ std::set<llvm::IRHash> oHashes;
+ std::vector<std::string> funcLocs;
+ Function *OrigFunc = nullptr;
+ for (const auto &FInfo : FInfos) {
+ OrigFunc = FInfo.F;
+
+ llvm::IRHash origHash = StructuralHash(*OrigFunc);
+ oHashes.insert(origHash);
+
+ // Print debug location.
+ std::string Result;
+ raw_string_ostream DbgLocOS(Result);
+ if (DISubprogram *DIS = OrigFunc->getSubprogram()) {
+ DebugLoc FuncDbgLoc =
+ DILocation::get(DIS->getContext(), DIS->getScopeLine(), 0, DIS);
+ FuncDbgLoc.print(DbgLocOS);
+ DbgLocOS.flush();
+ }
+ std::string singleLine =
+ "# functionLoc " +
+ std::to_string(GlobalValue::getGUID(OrigFunc->getName())) + " " +
+ Result + " " + std::string(OrigFunc->getName()) + "\n";
+ funcLocs.push_back(singleLine);
+ }
+}
+
+/// Merge all functions in \p FInfos by creating thunks which call the single
+/// merged function with additional parameters.
+void MergeFuncIgnoringConstImpl::mergeWithParams(const FunctionInfos &FInfos,
+ ParamInfos &Params) {
+ // We reuse the body of the first function for the new merged function.
+ Function *FirstF = FInfos.front().F;
+
+ // Build the type for the merged function. This will be the type of the
+ // original function (FirstF) but with the additional parameter which are
+ // needed to parameterize the merged function.
+ FunctionType *OrigTy = FirstF->getFunctionType();
+ SmallVector<Type *, 8> ParamTypes(OrigTy->param_begin(), OrigTy->param_end());
+
+ for (const ParamInfo &PI : Params) {
+ ParamTypes.push_back(PI.Values[0]->getType());
+ }
+
+ FunctionType *funcType =
+ FunctionType::get(OrigTy->getReturnType(), ParamTypes, false);
+
+ // Create the new function.
+ Function *NewFunction = Function::Create(funcType, FirstF->getLinkage(),
+ FirstF->getName() + ".Tm");
----------------
manman-ren wrote:
Addressed this in the stacked commit.
https://github.com/llvm/llvm-project/pull/68235
More information about the llvm-commits
mailing list