//===- Passes/GOTPass.cpp - Adds GOT entries ------------------------------===// // // The LLVM Linker // // This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// /// /// \file /// This linker pass transforms all GOT kind references to real references. /// That is, in assembly you can write something like: /// movq foo@GOTPCREL(%rip), %rax /// which means you want to load a pointer to "foo" out of the GOT (global /// Offsets Table). In the object file, the Atom containing this instruction /// has a Reference whose target is an Atom named "foo" and the Reference /// kind is a GOT load. The linker needs to instantiate a pointer sized /// GOT entry. This is done be creating a GOT Atom to represent that pointer /// sized data in this pass, and altering the Atom graph so the Reference now /// points to the GOT Atom entry (corresponding to "foo") and changing the /// Reference Kind to reflect it is now pointing to a GOT entry (rather /// then needing a GOT entry). /// /// There is one optimization the linker can do here. If the target of the GOT /// is in the same linkage unit and does not need to be interposable, and /// the GOT use is just a load (not some other operation), this pass can /// transform that load into an LEA (add). This optimizes away one memory load /// which at runtime that could stall the pipeline. This optimization only /// works for architectures in which a (GOT) load instruction can be change to /// an LEA instruction that is the same size. The method isGOTAccess() should /// only return true for "canBypassGOT" if this optimization is supported. /// //===----------------------------------------------------------------------===// #include "lld/Core/DefinedAtom.h" #include "lld/Core/File.h" #include "lld/Core/LLVM.h" #include "lld/Core/Pass.h" #include "lld/Core/Reference.h" #include "llvm/ADT/DenseMap.h" namespace lld { namespace { bool shouldReplaceTargetWithGOTAtom(const Atom *target, bool canBypassGOT) { // Accesses to shared library symbols must go through GOT. if (target->definition() == Atom::definitionSharedLibrary) return true; // Accesses to interposable symbols in same linkage unit must also go // through GOT. const DefinedAtom *defTarget = dyn_cast(target); if (defTarget != nullptr && defTarget->interposable() != DefinedAtom::interposeNo) { assert(defTarget->scope() != DefinedAtom::scopeTranslationUnit); return true; } // Target does not require indirection. So, if instruction allows GOT to be // by-passed, do that optimization and don't create GOT entry. return !canBypassGOT; } const DefinedAtom * findGOTAtom(const Atom *target, llvm::DenseMap &targetToGOT) { auto pos = targetToGOT.find(target); return (pos == targetToGOT.end()) ? nullptr : pos->second; } } // end anonymous namespace void GOTPass::perform(MutableFile &mergedFile) { // Use map so all pointers to same symbol use same GOT entry. llvm::DenseMap targetToGOT; // Scan all references in all atoms. for(const DefinedAtom *atom : mergedFile.defined()) { for (const Reference *ref : *atom) { // Look at instructions accessing the GOT. bool canBypassGOT; if (!isGOTAccess(ref->kind(), canBypassGOT)) continue; const Atom *target = ref->target(); assert(target != nullptr); if (!shouldReplaceTargetWithGOTAtom(target, canBypassGOT)) { // Update reference kind to reflect that target is a direct accesss. updateReferenceToGOT(ref, false); continue; } // Replace the target with a reference to a GOT entry. const DefinedAtom *gotEntry = findGOTAtom(target, targetToGOT); if (!gotEntry) { gotEntry = makeGOTEntry(*target); assert(gotEntry != nullptr); assert(gotEntry->contentType() == DefinedAtom::typeGOT); targetToGOT[target] = gotEntry; } const_cast(ref)->setTarget(gotEntry); // Update reference kind to reflect that target is now a GOT entry. updateReferenceToGOT(ref, true); } } // add all created GOT Atoms to master file for (auto &it : targetToGOT) { mergedFile.addAtom(*it.second); } } }