diff --git a/lld/Common/CommonLinkerContext.cpp b/lld/Common/CommonLinkerContext.cpp
--- a/lld/Common/CommonLinkerContext.cpp
+++ b/lld/Common/CommonLinkerContext.cpp
@@ -34,6 +34,12 @@
   // new in SpecificAlloc::create().
   for (auto &it : instances)
     it.second->~SpecificAllocBase();
+
+  for (PerThreadContext *context : perThreadContexts) {
+    for (auto &instance : context->perThreadInstances)
+      instance.second->~SpecificAllocBase();
+    delete context;
+  }
   lctx = nullptr;
 }
 
@@ -48,4 +54,5 @@
   if (lctx == nullptr)
     return;
   delete lctx;
+  lctx = nullptr;
 }
diff --git a/lld/Common/Memory.cpp b/lld/Common/Memory.cpp
--- a/lld/Common/Memory.cpp
+++ b/lld/Common/Memory.cpp
@@ -6,12 +6,35 @@
 //
 //===----------------------------------------------------------------------===//
 
-#include "lld/Common/Memory.h"
+#include <atomic>
+
 #include "lld/Common/CommonLinkerContext.h"
+#include "lld/Common/Memory.h"
 
 using namespace llvm;
 using namespace lld;
 
+namespace {
+// TODO: may need to move these to CommonLinkerContext class when we support
+// multiple processes.
+// Value of this is the index of the thread's Context in the
+// CommonLinkerContext.
+// XXX: 128-bit ints are not always lock-free, so limiting to 32-bit atomic for
+// now.
+thread_local uint32_t threadTag = 0;
+std::atomic<uint32_t> threadCounts(0);
+
+// Returns the index for this thread's context.
+inline uint32_t getThreadIndex() {
+  if (threadTag == 0) {
+    threadTag = ++threadCounts;
+  }
+  assert(threadTag < lld::MAX_THREADS);
+  return threadTag;
+}
+
+} // namespace
+
 SpecificAllocBase *
 lld::SpecificAllocBase::getOrCreate(void *tag, size_t size, size_t align,
                                     SpecificAllocBase *(&creator)(void *)) {
@@ -23,3 +46,26 @@
   }
   return instance;
 }
+
+PerThreadContext *CommonLinkerContext::perThreadContext() {
+  uint32_t threadIdx = getThreadIndex();
+  PerThreadContext *perThreadContext = perThreadContexts[threadIdx];
+  // Context didn't exist yet, so create a new one for this thread.
+  if (!perThreadContext)
+    perThreadContexts[threadIdx] = perThreadContext = new PerThreadContext;
+
+  return perThreadContext;
+}
+
+SpecificAllocBase *lld::SpecificAllocBase::getOrCreatePerThread(
+    void *tag, size_t size, size_t align,
+    SpecificAllocBase *(&creator)(void *)) {
+  PerThreadContext *threadContext = context().perThreadContext();
+  auto &instances = threadContext->perThreadInstances;
+  auto &instance = instances[tag];
+  if (instance == nullptr) {
+    void *storage = threadContext->bAlloc.Allocate(size, align);
+    instance = creator(storage);
+  }
+  return instance;
+}
diff --git a/lld/include/lld/Common/CommonLinkerContext.h b/lld/include/lld/Common/CommonLinkerContext.h
--- a/lld/include/lld/Common/CommonLinkerContext.h
+++ b/lld/include/lld/Common/CommonLinkerContext.h
@@ -22,6 +22,8 @@
 #include "lld/Common/ErrorHandler.h"
 #include "lld/Common/Memory.h"
 #include "llvm/Support/StringSaver.h"
+#include <array>
+#include <limits>
 
 namespace llvm {
 class raw_ostream;
@@ -29,6 +31,19 @@
 
 namespace lld {
 struct SpecificAllocBase;
+
+// Max number of threads ever created in a process.
+// Use int32's max value since that's the windows's limit.
+constexpr uint32_t MAX_THREADS = std::numeric_limits<int32>::max() - 1;
+
+// TODO Maybe just use the CommonLinkerContext class - but I dont want
+// unnecessary nesting of the mutexes
+struct PerThreadContext {
+  llvm::BumpPtrAllocator bAlloc;
+  llvm::DenseMap<void *, SpecificAllocBase *> perThreadInstances;
+  llvm::StringSaver perThreadSaver{bAlloc};
+};
+
 class CommonLinkerContext {
 public:
   CommonLinkerContext();
@@ -41,11 +56,19 @@
   llvm::DenseMap<void *, SpecificAllocBase *> instances;
 
   ErrorHandler e;
+
+  // Per thread allocs and savers.
+  PerThreadContext *perThreadContext();
+  // [0] is unused to make thread's index checking simpler, because 0 is
+  // sentinel for the index.
+  std::array<PerThreadContext *, MAX_THREADS> perThreadContexts;
 };
 
 // Retrieve the global state. Currently only one state can exist per process,
 // but in the future we plan on supporting an arbitrary number of LLD instances
 // in a single process.
+// TODO: How would this affect the TLS variable's addresses?
+// (hopefully not much)
 CommonLinkerContext &commonContext();
 
 template <typename T = CommonLinkerContext> T &context() {
@@ -54,8 +77,30 @@
 
 bool hasContext();
 
-inline llvm::StringSaver &saver() { return context().saver; }
-inline llvm::BumpPtrAllocator &bAlloc() { return context().bAlloc; }
+inline llvm::StringSaver &perThreadSaver() {
+  return context().perThreadContext()->perThreadSaver;
+}
+
+inline llvm::BumpPtrAllocator &perThreadbAlloc() {
+  return context().perThreadContext()->bAlloc;
+}
+
+inline llvm::StringSaver &saver() {
+#ifndef THREAD_SAFE_MEMORY
+  return context().perThreadContext()->perThreadSaver;
+#else
+  return context().saver;
+#endif
+}
+
+inline llvm::BumpPtrAllocator &bAlloc() {
+#ifndef THREAD_SAFE_MEMORY
+  return context().perThreadContext()->bAlloc;
+#else
+  return context().bAlloc;
+#endif
+}
+
 } // namespace lld
 
 #endif
diff --git a/lld/include/lld/Common/Memory.h b/lld/include/lld/Common/Memory.h
--- a/lld/include/lld/Common/Memory.h
+++ b/lld/include/lld/Common/Memory.h
@@ -22,14 +22,25 @@
 #define LLD_COMMON_MEMORY_H
 
 #include "llvm/Support/Allocator.h"
+#include "llvm/Support/RWMutex.h"
+
+// QUESTION: make this ON or OFF by default?
+#ifndef THREAD_SAFE_MEMORY
+#define THREAD_SAFE_MEMORY 1
+#endif
 
 namespace lld {
+
 // A base class only used by the CommonLinkerContext to keep track of the
 // SpecificAlloc<> instances.
 struct SpecificAllocBase {
   virtual ~SpecificAllocBase() = default;
   static SpecificAllocBase *getOrCreate(void *tag, size_t size, size_t align,
                                         SpecificAllocBase *(&creator)(void *));
+
+  static SpecificAllocBase *
+  getOrCreatePerThread(void *tag, size_t size, size_t align,
+                       SpecificAllocBase *(&creator)(void *));
 };
 
 // An arena of specific types T, created on-demand.
@@ -55,13 +66,38 @@
   return ((SpecificAlloc<T> *)instance)->alloc;
 }
 
+template <typename T>
+inline llvm::SpecificBumpPtrAllocator<T> &getSpecificAllocSingletonPerThread() {
+  SpecificAllocBase *instance = SpecificAllocBase::getOrCreatePerThread(
+      &SpecificAlloc<T>::tag, sizeof(SpecificAlloc<T>),
+      alignof(SpecificAlloc<T>), SpecificAlloc<T>::create);
+  return ((SpecificAlloc<T> *)instance)->alloc;
+}
 // Creates new instances of T off a (almost) contiguous arena/object pool. The
 // instances are destroyed whenever lldMain() goes out of scope.
-template <typename T, typename... U> T *make(U &&... args) {
+template <typename T, typename... U> T *makeUnsafe(U &&...args) {
   return new (getSpecificAllocSingleton<T>().Allocate())
       T(std::forward<U>(args)...);
 }
 
+// Creates new instances of T off a (almost) contiguous arena/object pool
+// specific for this thread. The instances are destroyed whenever lldMain() goes
+// out of scope.
+template <typename T, typename... U> T *makePerThread(U &&...args) {
+  return new (getSpecificAllocSingletonPerThread<T>().Allocate())
+      T(std::forward<U>(args)...);
+}
+
+template <typename T, typename... U> T *make(U &&...args) {
+#ifndef THREAD_SAFE_MEMORY
+  return new (getSpecificAllocSingleton<T>().Allocate())
+      T(std::forward<U>(args)...);
+#else
+  return new (getSpecificAllocSingletonPerThread<T>().Allocate())
+      T(std::forward<U>(args)...);
+#endif
+}
+
 } // namespace lld
 
 #endif