[libcxx-commits] [libcxx] [libc++] Simplify the __libcpp_refstring implementation (PR #166344)

[Nikolas Klauser via libcxx-commits]

https://github.com/philnik777 created https://github.com/llvm/llvm-project/pull/166344

We don't support any platforms anymore where the `__libcpp_refstring` has to be ABI compatible with libstdc++'s C++03 implementation of `string`, so we can significantly simplify the implementation.


>From a43c3dfbf896e1f9b6efd4d87c21cfdc9df05c6c Mon Sep 17 00:00:00 2001
From: Nikolas Klauser <nikolasklauser at berlin.de>
Date: Tue, 4 Nov 2025 11:15:59 +0100
Subject: [PATCH] [libc++] Simplify the __libcpp_refstring implementation

---
 libcxx/include/stdexcept       |   7 +--
 libcxx/src/include/refstring.h | 103 ++++++---------------------------
 2 files changed, 20 insertions(+), 90 deletions(-)

diff --git a/libcxx/include/stdexcept b/libcxx/include/stdexcept
index 85e11629bd6e3..8082c4dc8b0fe 100644
--- a/libcxx/include/stdexcept
+++ b/libcxx/include/stdexcept
@@ -57,9 +57,8 @@ _LIBCPP_BEGIN_NAMESPACE_STD
 
 #  ifndef _LIBCPP_ABI_VCRUNTIME
 class _LIBCPP_HIDDEN __libcpp_refstring {
-  const char* __imp_;
-
-  bool __uses_refcount() const;
+  struct __rep;
+  __rep* __imp_;
 
 public:
   explicit __libcpp_refstring(const char* __msg);
@@ -67,7 +66,7 @@ public:
   __libcpp_refstring& operator=(const __libcpp_refstring& __s) _NOEXCEPT;
   ~__libcpp_refstring();
 
-  _LIBCPP_HIDE_FROM_ABI const char* c_str() const _NOEXCEPT { return __imp_; }
+  _LIBCPP_HIDE_FROM_ABI const char* c_str() const _NOEXCEPT;
 };
 #  endif // !_LIBCPP_ABI_VCRUNTIME
 
diff --git a/libcxx/src/include/refstring.h b/libcxx/src/include/refstring.h
index 3e0ec7a97c7be..40ae4dafb6ae6 100644
--- a/libcxx/src/include/refstring.h
+++ b/libcxx/src/include/refstring.h
@@ -15,112 +15,43 @@
 #include <cstring>
 #include <stdexcept>
 
-// MacOS and iOS used to ship with libstdc++, and still support old applications
-// linking against libstdc++. The libc++ and libstdc++ exceptions are supposed
-// to be ABI compatible, such that they can be thrown from one library and caught
-// in the other.
-//
-// For that reason, we must look for libstdc++ in the same process and if found,
-// check the string stored in the exception object to see if it is the GCC empty
-// string singleton before manipulating the reference count. This is done so that
-// if an exception is created with a zero-length string in libstdc++, libc++abi
-// won't try to delete the memory.
-#if defined(__ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__) || defined(__ENVIRONMENT_IPHONE_OS_VERSION_MIN_REQUIRED__)
-#  define _LIBCPP_CHECK_FOR_GCC_EMPTY_STRING_STORAGE
-#  include <dlfcn.h>
-#  include <mach-o/dyld.h>
-#endif
-
 _LIBCPP_BEGIN_NAMESPACE_STD
 
-namespace __refstring_imp {
-namespace {
-typedef int count_t;
+_LIBCPP_CLANG_DIAGNOSTIC_IGNORED("-Wc99-extensions")
 
-struct _Rep_base {
-  std::size_t len;
-  std::size_t cap;
-  count_t count;
+struct __libcpp_refstring::__rep {
+  ptrdiff_t count;
+  char data[];
 };
 
-inline _Rep_base* rep_from_data(const char* data_) noexcept {
-  char* data = const_cast<char*>(data_);
-  return reinterpret_cast<_Rep_base*>(data - sizeof(_Rep_base));
-}
-
-inline char* data_from_rep(_Rep_base* rep) noexcept {
-  char* data = reinterpret_cast<char*>(rep);
-  return data + sizeof(*rep);
-}
-
-#if defined(_LIBCPP_CHECK_FOR_GCC_EMPTY_STRING_STORAGE)
-inline const char* compute_gcc_empty_string_storage() noexcept {
-  void* handle = dlopen("/usr/lib/libstdc++.6.dylib", RTLD_NOLOAD);
-  if (handle == nullptr)
-    return nullptr;
-  void* sym = dlsym(handle, "_ZNSs4_Rep20_S_empty_rep_storageE");
-  if (sym == nullptr)
-    return nullptr;
-  return data_from_rep(reinterpret_cast<_Rep_base*>(sym));
-}
-
-inline const char* get_gcc_empty_string_storage() noexcept {
-  static const char* p = compute_gcc_empty_string_storage();
-  return p;
-}
-#endif
-
-} // namespace
-} // namespace __refstring_imp
-
-using namespace __refstring_imp;
-
 inline __libcpp_refstring::__libcpp_refstring(const char* msg) {
   std::size_t len = strlen(msg);
-  _Rep_base* rep  = static_cast<_Rep_base*>(::operator new(sizeof(*rep) + len + 1));
-  rep->len        = len;
-  rep->cap        = len;
+  auto* rep       = static_cast<__rep*>(::operator new(sizeof(__rep) + len + 1));
   rep->count      = 0;
-  char* data      = data_from_rep(rep);
-  std::memcpy(data, msg, len + 1);
-  __imp_ = data;
+  std::memcpy(rep->data, msg, len + 1);
+  __imp_ = rep;
 }
 
 inline __libcpp_refstring::__libcpp_refstring(const __libcpp_refstring& s) noexcept : __imp_(s.__imp_) {
-  if (__uses_refcount())
-    __libcpp_atomic_add(&rep_from_data(__imp_)->count, 1);
+  __libcpp_atomic_add(&__imp_->count, 1);
 }
 
 inline __libcpp_refstring& __libcpp_refstring::operator=(__libcpp_refstring const& s) noexcept {
-  bool adjust_old_count     = __uses_refcount();
-  struct _Rep_base* old_rep = rep_from_data(__imp_);
-  __imp_                    = s.__imp_;
-  if (__uses_refcount())
-    __libcpp_atomic_add(&rep_from_data(__imp_)->count, 1);
-  if (adjust_old_count) {
-    if (__libcpp_atomic_add(&old_rep->count, count_t(-1)) < 0) {
-      ::operator delete(old_rep);
-    }
-  }
+  __rep* old_rep = __imp_;
+  __imp_         = s.__imp_;
+  __libcpp_atomic_add(&__imp_->count, 1);
+
+  if (__libcpp_atomic_add(&old_rep->count, ptrdiff_t(-1)) < 0)
+    ::operator delete(old_rep);
   return *this;
 }
 
 inline __libcpp_refstring::~__libcpp_refstring() {
-  if (__uses_refcount()) {
-    _Rep_base* rep = rep_from_data(__imp_);
-    if (__libcpp_atomic_add(&rep->count, count_t(-1)) < 0) {
-      ::operator delete(rep);
-    }
-  }
+  if (__libcpp_atomic_add(&__imp_->count, ptrdiff_t(-1)) < 0)
+    ::operator delete(__imp_);
 }
 
-inline bool __libcpp_refstring::__uses_refcount() const {
-#if defined(_LIBCPP_CHECK_FOR_GCC_EMPTY_STRING_STORAGE)
-  return __imp_ != get_gcc_empty_string_storage();
-#else
-  return true;
-#endif
-}
+inline const char* __libcpp_refstring::c_str() const noexcept { return __imp_->data; }
 
 _LIBCPP_END_NAMESPACE_STD