[libc++] Use ABI tags instead of internal linkage to provide per-TU insulation

[ldionne]

Summary

Instead of marking private symbols with internal_linkage (which leads to
one copy per translation unit -- rather wasteful), use an ABI tag that
gets rev'd with each libc++ version. That way, we know that we can't have
name collisions between implementation-detail functions across libc++
versions, so we'll never violate the ODR. However, within a single program,
each symbol still has a proper name with external linkage, which means
that the linker is free to deduplicate symbols even across TUs.

This actually means that we can guarantee that versions of libc++ can
be mixed within the same program without ever having to take a code size
hit, and without having to manually opt-in -- it should just work out of
the box.

[ldionne]

It would be awesome if someone with deeper linker expertise could double-check this change. I think it should work in principle, but it would be nice to know if there are pitfalls I'm missing.

[ldionne]

I dicussed this internally with linker folks and they confirmed that it should work. I'll wait a bit before landing this, but I think we've got something. @EricWF it would be nice to get your eyes on this as well!

[EricWF]

Hmm, so I'm not sure this is the best idea. At worst, I think it'll introduce ODR bugs if we use it on anything other than inline functions [2], and when applied only to inline functions it helps only in a very very limited number of cases [1].

[1]. Since most changes to a function signature affect the mangling as well, there's no additional benefit from further changing the mangling using __abi_tag__. The one case it would catch is if we change only the return type of a non-template function (since the return type is mangled into function templates).
And we shouldn't be changing the return type of user-facing functions, like ever.

[2] All of the following examples break when the attribute is applied to a class type.

// RUN: clang++ -c -o test1.o -DTAG=1 test.cpp && clang++ -o a.out -DTAG=2 test.cpp test1.o && ./a.out
#include <typeinfo>
#include <cassert>
#include <cstdio>

#ifndef TAG
#error TAG must be defined
#endif
#define TO_STR2(X) #X
#define TO_STR(X) TO_STR2(X)

#define HIDDEN __attribute__((__abi_tag__(TO_STR(TAG))))

struct HIDDEN  MyType {
  virtual ~MyType() = default;
  static  int data;
};

inline  int MyType::data = 42;

int* get_data();
const std::type_info* get_info();
void throw_type();

__attribute__((weak)) void passes_param(MyType); // weak only used to demonstrate linker failure at runtime rather than linktime.

#if TAG == 1
int* get_data() {
  return &MyType::data;
}
void throw_type() {
  MyType t{};
  throw t;
}

const std::type_info* get_info() {
  return &typeid(MyType);
}
void passes_param(MyType) {
  printf("INFO: Passed to MyType@TAG=1");
}
#else
int main() {
  int num_errors = 0;
  if (&passes_param) {
    printf("SUCCESS: passes_param(MyType) has definition\n");
  } else {
    ++num_errors;
    printf("ERROR: passes_param(MyType) is undefined\n");
  }
  if (get_data() == &MyType::data) {
    printf("SUCCESS: data@TAG=1 == data@TAG=2\n");
  } else {

    printf("ERROR: data@TAG=1 != data@TAG=2\n");
    ++num_errors;
  }
  if (*get_info() == typeid(MyType)) {

    printf("typeid(MyType)@TAG=1 == typeid(MyType)@TAG=2\n");
  } else {
    printf("ERROR: typeid(MyType)@TAG=1 == typeid(MyType)@TAG=2\n");
    ++num_errors ;
  }
  try {
    throw_type();
    assert(false);
  } catch (MyType const& t) {
    printf("INFO: exception MyType@TAG=1 caught as MyType@TAG=2\n");
  } catch (...) {
    printf("ERROR: exception MyType@TAG=1 not caught\n");
    ++num_errors;
  }
  printf("%d errors...\n", num_errors);
  return num_errors;
}
#endif

libcxx/include/__config

677

Symbols that are linkonce_odr and that get emitted inside a users dylib are never linked to from outside the dylib. So they're already hermetic in that way.

681

This actually seems like a rather big footgun to me. At least if it's ever applied to anything other than an inline function.

In the case where it's applied to either (A) a type, or (B) a data member, then, Instead of ensuring that no ODR violations arise, this all but ensures ODR violations do arise. As a result any behavior that depends on uniqueness (typeid, exceptions, static data members, vtables, type mangling, anything with a unique address) will now have one definition per libc++ version, and valid behavior break.

[ldionne]

In D127444#3578601, @EricWF wrote:

Hmm, so I'm not sure this is the best idea. At worst, I think it'll introduce ODR bugs if we use it on anything other than inline functions [2], and when applied only to inline functions it helps only in a very very limited number of cases [1].

Thanks a lot for taking a look. I think it's important to understand that _LIBCPP_HIDE_FROM_ABI is only (or at least should only) ever be applied to inline functions. That's required for it today already, since we use the internal_linkage attribute. In fact, all of the same failures that you rightly point out in your example already fail today when we use the internal_linkage attribute, because each TU will get its own definition of the RTTI/function-local-static/whatever.

Furthermore, if someone mixes two different versions of libc++ in the same final linked image without using internal_linkage today (aka without specifying _LIBCPP_HIDE_FROM_ABI_PER_TU=1), then they are at risk of ODR violations because we don't know which version of (for example) __copy_impl the linker will use. This applies to basically any internal function that we permit ourselves to change the preconditions/postconditions of without changing the name or the mangling, which is effectively anything that's marked as _LIBCPP_HIDE_FROM_ABI.

So, in other words -- you are correct that the cases that you highlight will fail with this change, but they already do. And in the configuration where these tests wouldn't fail, users are not supposed to be mixing versions of libc++. With that in mind, do you think this change makes more sense?

libcxx/include/__config

677

I mean, they can still get exported from the dylib if they have default visibility, no?

They generally won't be used because anyone compiling against the same headers would *also* get a linkonce_odr definition in their own dylib/exe and their code would use that, but technically, the symbols are still exported from (both) final linked images. Unless I missed something special about linkonce_odr symbols?

[EricWF]

OK, so i'm convinced this change should be OK assuming we continue to ensure _LIBCPP_HIDE_FROM_ABI is only ever applied to inline functions.
It would also need to be applied to every single inline function the library produces, otherwise a tagged function could call an untagged function,
since the untagged definition can be selected at random from the available definitions.

The other consideration is the cost of this change. While small, it does increase the mangling of every single inline function in libc++ by ~5 bytes.
After a little testing, this appears to translate into an object file size increase of ~1% when debug information is enabled.

My current opinion is that the binary size increase outweighs the limited benefits the change provides.
But I'm open to being convinced.

@ldionne maybe you could produce a compiling example of a situation this change intends to fix?

libcxx/include/__config

677

You're right. The linker will choose the symbol in an SO if it absolutely must. But that's only if we don't mark it with __attribute__((visibility("hidden")), which we should be using, since it's both necessary and sufficient to solve the export problem.

The ABI tag alone does nothing to prevent symbol exporting. For example assuming binary A gets symbol X@V1 from libB.so, then libB.so is recompiled against a new version, so it exports X@V2 instead now. Binary A breaks all the same.

It's also worth noting that we're assuming the ABI has not changed apart from inline functions potentially, otherwise the ABI namespace should have changed, solving the ODR problem entirely.

Which limits the utility of this to cases where

1. a user *statically links* object files produced against two different versions of libc++, AND:
   1. a non-template inline function F changes its return type, without changing anything else, so that the mangling stays the same, but the ABI changes; Or,
   2. A inline function G changes it's pre-conditions/post-conditions without changing the signature, so that input that's valid to G_0 is not valid to G_1 or vice-versa.

Arguably, the user shouldn't be doing (1), and we shouldn't make changes of the kind required for (A) or (B) (that don't also change mangling).
If a user does to (1), but we're careful to not do (A) or (B), then we have a no ODR violation at all -- because either:

1. all the available definitions of a function F are the same, or
2. we have a "benign" ODR violation where there are multiple different definitions of F, but selecting either definition is OK because the pre/post conditions haven't changed.

[philnik]

Just out of interest: Does this remove the problem that we always_inline a lot of functions with GCC?

[EricWF]

I should add...

We should be explicitly aware of changes that affect the pre/post conditions of a libc++ function, and only do so consciously;
Allowing us to explicitly change the mangling using __abi_tag__ or any other means (typically such changes already rename the function or change its signature)

[EricWF]

In D127444#3579756, @philnik wrote:

Just out of interest: Does this remove the problem that we always_inline a lot of functions with GCC?

At this point I hope we're only always_inlining functions for performance/code size reasons, rather than
as ABI management. Though it may still be needed to battle with extern template semantics.

Specifically which always inline usages are you talking about?

[philnik]

In D127444#3579773, @EricWF wrote:

In D127444#3579756, @philnik wrote:

Just out of interest: Does this remove the problem that we always_inline a lot of functions with GCC?

At this point I hope we're only always_inlining functions for performance/code size reasons, rather than
as ABI management. Though it may still be needed to battle with extern template semantics.

Specifically which always inline usages are you talking about?

I mean that GCC doesn't provide internal_linkage (https://godbolt.org/z/8jTzq8zvz), which means that we fall back to always_inline with GCC with _LIBCPP_HIDE_FROM_ABI_PER_TU enabled.

[philnik]

Oh and apparently also without it. We define _LIBCPP_EXCLUDE_FROM_EXPLICIT_INSTANTIATION as always_inline if it's not provided (which it isn't with GCC: https://godbolt.org/z/5vvr7ne94).

[ldionne]

(consolidating here in a single thread)

Edit: I'm using the term "implementation detail function" below, but I really mean "functions that we don't want to appear as part of our ABI", aka all functions that are marked with _LIBCPP_HIDE_FROM_ABI or _LIBCPP_INLINE_VISIBILITY today.

In D127444#3579773, @EricWF wrote:

You're right. The linker will choose the symbol in an SO if it absolutely must. But that's only if we don't mark it with __attribute__((visibility("hidden")), which we should be using, since it's both necessary and sufficient to solve the export problem.

We agree here. We do mark implementation-detail symbols with visibility("hidden") for that purpose (via _LIBCPP_HIDDEN).

Which limits the utility of this to cases where

1. a user statically links object files produced against two different versions of libc++, AND:
   
   A. a non-template inline function F changes its return type, without changing anything else, so that the mangling stays the same, but the ABI changes; Or,
   
   B. a inline function G changes it's pre-conditions/post-conditions without changing the signature, so that input that's valid to G_0 is not valid to G_1 or vice-versa.

We're in agreement here. Solving this problem has been the goal of this whole endeavour with internal_linkage and always_inline from the beginning. Initially, libc++ marked everything with always_inline so as to "avoid" this problem entirely (if you don't emit symbols, you can't have ODR issues!). This was true until I made changes to start using internal_linkage circa 2018. The goal of that change was to stop inlining everything, which was bad for code size and debugging.

Arguably, the user shouldn't be doing (1), and we shouldn't make changes of the kind required for (A) or (B) (that don't also change mangling).
If a user does to (1), but we're careful to not do (A) or (B), then we have a no ODR violation at all -- because either:

1. all the available definitions of a function F are the same, or
2. we have a "benign" ODR violation where there are multiple different definitions of F, but selecting either definition is OK because the pre/post conditions haven't changed.

We're in partial agreement. The ability for users to mix statically linked versions of libc++ is something that we used to support, until I changed it to be conditionally supported by vendors (based on whether LIBCXX_HIDE_FROM_ABI_PER_TU_BY_DEFAULT is used when configuring the library). We still support it on Apple platforms, and AFAICT some users might rely on it.

Where we disagree is when it comes to what is reasonable from libc++ to avoid these ODR violations or make them benign. We currently don't pay attention to how we change implementation-detail functions, and I don't think it is realistic to do it robustly. Think about it -- it means that we would need to know the full set of function signatures (name + parameter list) to have ever been used in a shipping libc++, and we should either never reuse any of these signatures, or if we do, then make sure we have compatible pre/post conditions for that function. For example, can we guarantee that we haven't previously shipped a function named __copy_impl(char*, char*, char*) with pre/post conditions incompatible with the ones we have in our code today? Even back in e.g. LLVM 7?

As you mention, the only way to do it robustly is to never change any function but in minor ways that would make an ODR mismatch effectively a benign ODR violation. However, that places a significant restriction on how we can evolve libc++'s code, and it also increases the barrier to contributing for reasons intangible to anyone not having read this discussion. Since we have a clear technical solution instead, I think we should go for it.

In D127444#3579745, @EricWF wrote:

OK, so i'm convinced this change should be OK assuming we continue to ensure _LIBCPP_HIDE_FROM_ABI is only ever applied to inline functions.

Agreed, that's a requirement, and it was also a requirement prior to this change -- we never use _LIBCPP_HIDE_FROM_ABI on non-functions.

It would also need to be applied to every single inline function the library produces, otherwise a tagged function could call an untagged function,
since the untagged definition can be selected at random from the available definitions.

Agreed, and we already do that except:

1. For a few functions that have been forgotten (we sometimes catch missing _LIBCPP_HIDE_FROM_ABI and we add them), and
2. for a few functions that were large enough that we didn't want to mark with internal_linkage or always_inline, and for which we do go through the trouble of making sure we don't introduce ODR violations. That being said, with this proposed patch, we could mark those functions as well without problem.

The other consideration is the cost of this change. While small, it does increase the mangling of every single inline function in libc++ by ~5 bytes.
After a little testing, this appears to translate into an object file size increase of ~1% when debug information is enabled.

My current opinion is that the binary size increase outweighs the limited benefits the change provides.
But I'm open to being convinced.

I actually think that's the crux of this change.

1. For configurations where we use LIBCXX_HIDE_FROM_ABI_PER_TU_BY_DEFAULT=ON, the benefit will be that we stop having duplicate instantiations of things like pop_heap in each translation unit where it is used. This is just a random one, but all functions with _LIBCPP_HIDE_FROM_ABI will get that benefit. For this configuration, this is a net win, there is no cost I can think of.
2. For configurations of the library where LIBCXX_HIDE_FROM_ABI_PER_TU_BY_DEFAULT=OFF (e.g. the LLVM release), this will provide the ability for users to statically link against mixed versions of libc++ safely (which some users probably do without even knowing it, but is currently unsafe). The cost will be the ABI tag, which is roughly 5 bytes per symbol, as you mention.

For (2), the cost should only affect the size of object files, since these symbols should be stripped from shipping executables. Given that, my inclination would be to raise correctness above object size, and to try to push efforts such as ABI compression mangling instead, which could deliver absolutely insane improvements to symbol sizes, much beyond saving 5 bytes.

@ldionne maybe you could produce a compiling example of a situation this change intends to fix?

Like I said, this is not really to fix any single specific situation -- it's to (1) reduce per-TU duplication when per-TU insulation is provided, (2) provide an ABI safety guarantee to users in all cases, and (3) simplify the library.

Another thing I can do is keep the LIBCXX_HIDE_FROM_ABI_PER_TU_BY_DEFAULT switch and not provide the ABI safety guarantee for vendors who don't care about it (that way we don't have the ABI tag when we previously wouldn't have had internal_linkage). However, this keeps existing complexity around, and to quote someone I respect a lot: "most users don't know what their ABI needs are, and we need to pick what's right for them". IMO this is such a case.

In D127444#3579773, @EricWF wrote:

In D127444#3579756, @philnik wrote:

Just out of interest: Does this remove the problem that we always_inline a lot of functions with GCC?

At this point I hope we're only always_inlining functions for performance/code size reasons, rather than
as ABI management. Though it may still be needed to battle with extern template semantics.

Specifically which always inline usages are you talking about?

It does not entirely remove the need for always_inline, because GCC doesn't support exclude_from_explicit_instantiation. However, it does take us one step closer to not needing always_inline for ABI management. In particular, I plan to do a follow-up patch that would create an inline namespace that gets rev'd at each release for implementation-detail functions, and that would remove the need to use _LIBCPP_HIDE_FROM_ABI on declarations in that namespace (and that would allow ditching always_inline for many functions on GCC).

[ldionne]

Rebase onto main. Gentle ping.

[EricWF]

It's true that this is a strict improvement over the internal_I guess I'm fine with this if we give Google a way to turn it off, so we don't have to pay the cost of the larger manglings.

libcxx/include/__config

637

Can we give users an ability to turn this off?

libcxx/test/libcxx/strings/basic.string/PR42676.sh.cpp

7

This test may still have value if we change it to exercise the __abi_tag__ path instead?

[EricWF]

Actually, I think we need to investigate the mangling scheme when the provided ABI tag is a number, because the mangling is B<size-of-tag><tag>, so for example the ABI tag "abc" on foo produces "_Z3fooB3abcv", where 3 is the length of the tag. Now consider the tag 1, which produces B11v, which looks as though it's a tag of length 11.
This seems to break some demanglers, maybe it also will break more?

[EricWF]

Ah, I see we limit ourselves to valid identifiers. LGTM again, just with a knob to turn it off.

[ldionne]

In D127444#3609719, @EricWF wrote:

Ah, I see we limit ourselves to valid identifiers. LGTM again, just with a knob to turn it off.

I'm against this on principle, but I'll do it to unblock this patch. IMO this is the wrong way to reduce the size of object files -- we should be using compression mangling instead, and it's not a super idea to change our overall ABI stability promise just for that reason. But as I said, I'll add an escape hatch with a comment about using compression mangling instead.

[ldionne]

Add escape hatch

[JDevlieghere]

Hey Louis,

This change breaks two LLDB tests: TestQueueFromStdModule.py and TestStackFromStdModule.py. I'm not at all familiar with these tests, but it seems the error mode is that the debugger can't find certain symbols:

error: Couldn't lookup symbols:
  __ZNSt3__1miERKNS_13move_iteratorIPP1CEES6_

Given that the bot has been red for 20 hours I went ahead and reverted your change in 61d417ceff90c4bd99bfed082d81bcc33ecf5a45.

[JDevlieghere]

Link to the failures:

https://green.lab.llvm.org/green/view/LLDB/job/lldb-cmake/45130/testReport/lldb-api/commands_expression_import-std-module_stack/TestStackFromStdModule_py/
https://green.lab.llvm.org/green/view/LLDB/job/lldb-cmake/45130/testReport/lldb-api/commands_expression_import-std-module_queue/TestQueueFromStdModule_py/

[thakis]

When this relands, maybe you could reland https://github.com/llvm/llvm-project/commit/6148c79a64ff03e5876882dd90d895ddb1783778 with it. (If not, no worries.)

[ldionne]

Band-aid for the LLDB issue, and also include 6148c79a in the commit.

libcxx/CMakeLists.txt

set(LIBCXX_TYPEINFO_COMPARISON_IMPLEMENTATION "default" CACHE STRING
    See the documentation in <libcxx/include/typeinfo> for details on what each
    value means.")
 set(TYPEINFO_COMPARISON_VALUES "default;1;2;3")
 if (NOT ("${LIBCXX_TYPEINFO_COMPARISON_IMPLEMENTATION}" IN_LIST TYPEINFO_COMPARISON_VALUES))
   message(FATAL_ERROR "Value '${LIBCXX_TYPEINFO_COMPARISON_IMPLEMENTATION}' is not a valid value for
                        LIBCXX_TYPEINFO_COMPARISON_IMPLEMENTATION")
 endif()
 
-option(LIBCXX_HIDE_FROM_ABI_PER_TU_BY_DEFAULT "Enable per TU ABI insulation by default. To be used by vendors." OFF)
 set(LIBCXX_ABI_DEFINES "" CACHE STRING "A semicolon separated list of ABI macros to define in the site config header.")
 option(LIBCXX_EXTRA_SITE_DEFINES "Extra defines to add into __config_site")
 option(LIBCXX_USE_COMPILER_RT "Use compiler-rt instead of libgcc" OFF)
 
 cmake_dependent_option(LIBCXX_ENABLE_BACKWARDS_COMPATIBILITY_DEBUG_MODE_SYMBOLS
   "Whether to include the old Debug mode symbols in the compiled library. This
    is provided for backwards compatibility since the compiled library used to
    always contain those symbols, regardless of whether the library was built
function(cxx_add_windows_flags target)
   endif()
 endfunction()
 
 # Configuration file flags =====================================================
 config_define(${LIBCXX_ABI_VERSION} _LIBCPP_ABI_VERSION)
 config_define(${LIBCXX_ABI_NAMESPACE} _LIBCPP_ABI_NAMESPACE)
 config_define_if(LIBCXX_ABI_FORCE_ITANIUM _LIBCPP_ABI_FORCE_ITANIUM)
 config_define_if(LIBCXX_ABI_FORCE_MICROSOFT _LIBCPP_ABI_FORCE_MICROSOFT)
-config_define_if(LIBCXX_HIDE_FROM_ABI_PER_TU_BY_DEFAULT _LIBCPP_HIDE_FROM_ABI_PER_TU_BY_DEFAULT)
 config_define_if_not(LIBCXX_ENABLE_THREADS _LIBCPP_HAS_NO_THREADS)
 config_define_if_not(LIBCXX_ENABLE_MONOTONIC_CLOCK _LIBCPP_HAS_NO_MONOTONIC_CLOCK)
 if (NOT LIBCXX_TYPEINFO_COMPARISON_IMPLEMENTATION STREQUAL "default")
   config_define("${LIBCXX_TYPEINFO_COMPARISON_IMPLEMENTATION}" _LIBCPP_TYPEINFO_COMPARISON_IMPLEMENTATION)
 endif()
 config_define_if(LIBCXX_HAS_PTHREAD_API _LIBCPP_HAS_THREAD_API_PTHREAD)
 config_define_if(LIBCXX_HAS_EXTERNAL_THREAD_API _LIBCPP_HAS_THREAD_API_EXTERNAL)
 config_define_if(LIBCXX_HAS_WIN32_THREAD_API _LIBCPP_HAS_THREAD_API_WIN32)

libcxx/cmake/caches/Apple.cmake

 set(CMAKE_BUILD_TYPE MinSizeRel CACHE STRING "")
 set(CMAKE_POSITION_INDEPENDENT_CODE OFF CACHE BOOL "")
 
 set(LIBCXX_USE_COMPILER_RT ON CACHE BOOL "")
 set(LIBCXX_ENABLE_ASSERTIONS OFF CACHE BOOL "")
 set(LIBCXX_ABI_VERSION "1" CACHE STRING "")
 set(LIBCXX_ENABLE_EXPERIMENTAL_LIBRARY OFF CACHE BOOL "")
 set(LIBCXX_ENABLE_STATIC ON CACHE BOOL "")
 set(LIBCXX_ENABLE_SHARED ON CACHE BOOL "")
 set(LIBCXX_CXX_ABI libcxxabi CACHE STRING "")
-set(LIBCXX_HIDE_FROM_ABI_PER_TU_BY_DEFAULT ON CACHE BOOL "")
 set(LIBCXX_ENABLE_BACKWARDS_COMPATIBILITY_DEBUG_MODE_SYMBOLS OFF CACHE BOOL "")
 set(LIBCXX_ENABLE_VENDOR_AVAILABILITY_ANNOTATIONS ON CACHE BOOL "")
 set(LIBCXX_ENABLE_INCOMPLETE_FEATURES OFF CACHE BOOL "")
 
 set(LIBCXX_HERMETIC_STATIC_LIBRARY ON CACHE BOOL "")
 set(LIBCXXABI_HERMETIC_STATIC_LIBRARY ON CACHE BOOL "")
 
 set(LIBCXXABI_ENABLE_ASSERTIONS OFF CACHE BOOL "")
 set(LIBCXXABI_ENABLE_FORGIVING_DYNAMIC_CAST ON CACHE BOOL "")
 
 set(LIBCXX_TEST_CONFIG "apple-libc++-shared.cfg.in" CACHE STRING "")
 set(LIBCXXABI_TEST_CONFIG "apple-libc++abi-shared.cfg.in" CACHE STRING "")
 set(LIBCXX_TEST_PARAMS "stdlib=apple-libc++" CACHE STRING "")
 set(LIBCXXABI_TEST_PARAMS "${LIBCXX_TEST_PARAMS}" CACHE STRING "")

libcxx/docs/BuildingLibcxx.rst

.. option:: LIBCXX_BENCHMARK_NATIVE_STDLIB:STRING
   compare libc++ to libstdc++ by building the benchmark tests against both
   standard libraries.
 
 .. option:: LIBCXX_BENCHMARK_NATIVE_GCC_TOOLCHAIN:STRING
 
   Use the specified GCC toolchain and standard library when building the native
   stdlib benchmark tests.
 
-.. option:: LIBCXX_HIDE_FROM_ABI_PER_TU_BY_DEFAULT:BOOL
-
-  **Default**: ``OFF``
-
-  Pick the default for whether to constrain ABI-unstable symbols to
-  each individual translation unit. This setting controls whether
-  `_LIBCPP_HIDE_FROM_ABI_PER_TU_BY_DEFAULT` is defined by default --
-  see the documentation of that macro for details.
-
 
 libc++ ABI Feature Options
 --------------------------
 
 The following options allow building libc++ for a different ABI version.
 
 .. option:: LIBCXX_ABI_VERSION:STRING
 

libcxx/docs/DesignDocs/VisibilityMacros.rst

**_LIBCPP_HIDE_FROM_ABI_AFTER_V1**
   from the ABI (if the symbol is not part of the ABI in the version we're
   building), or we leave it included. From user code (i.e. when we're not
   building libc++), the macro always marks symbols as internal so that programs
   built using new libc++ headers stop relying on symbols that are removed from
   the ABI in a future version. Each time we release a new stable version of the
   ABI, we should create a new _LIBCPP_HIDE_FROM_ABI_AFTER_XXX macro, and we can
   use it to start removing symbols from the ABI after that stable version.
 
-**_LIBCPP_HIDE_FROM_ABI_PER_TU**
-  This macro controls whether symbols hidden from the ABI with `_LIBCPP_HIDE_FROM_ABI`
-  are local to each translation unit in addition to being local to each final
-  linked image. This macro is defined to either 0 or 1. When it is defined to
-  1, translation units compiled with different versions of libc++ can be linked
-  together, since all non ABI-facing functions are local to each translation unit.
-  This allows static archives built with different versions of libc++ to be linked
-  together. This also means that functions marked with `_LIBCPP_HIDE_FROM_ABI`
-  are not guaranteed to have the same address across translation unit boundaries.
-
-  When the macro is defined to 0, there is no guarantee that translation units
-  compiled with different versions of libc++ can interoperate. However, this
-  leads to code size improvements, since non ABI-facing functions can be
-  deduplicated across translation unit boundaries.
-
-  This macro can be defined by users to control the behavior they want from
-  libc++. The default value of this macro (0 or 1) is controlled by whether
-  `_LIBCPP_HIDE_FROM_ABI_PER_TU_BY_DEFAULT` is defined, which is intended to
-  be used by vendors only (see below).
-
-**_LIBCPP_HIDE_FROM_ABI_PER_TU_BY_DEFAULT**
-  This macro controls the default value for `_LIBCPP_HIDE_FROM_ABI_PER_TU`.
-  When the macro is defined, per TU ABI insulation is enabled by default, and
-  `_LIBCPP_HIDE_FROM_ABI_PER_TU` is defined to 1 unless overridden by users.
-  Otherwise, per TU ABI insulation is disabled by default, and
-  `_LIBCPP_HIDE_FROM_ABI_PER_TU` is defined to 0 unless overridden by users.
-
-  This macro is intended for vendors to control whether they want to ship
-  libc++ with per TU ABI insulation enabled by default. Users can always
-  control the behavior they want by defining `_LIBCPP_HIDE_FROM_ABI_PER_TU`
-  appropriately.
-
-  By default, this macro is not defined, which means that per TU ABI insulation
-  is not provided unless explicitly overridden by users.
-
 **_LIBCPP_TYPE_VIS**
   Mark a type's typeinfo, vtable and members as having default visibility.
   This attribute cannot be used on class templates.
 
 **_LIBCPP_TEMPLATE_VIS**
   Mark a type's typeinfo and vtable as having default visibility.
   This macro has no effect on the visibility of the type's member functions.
 

 
 **_LIBCPP_EXCEPTION_ABI**
   Mark the member functions, typeinfo, and vtable of the type as being exported
   by the libc++ library. This macro must be applied to all *exception types*.
   Exception types should be defined directly in namespace `std` and not the
   versioning namespace. This allows throwing and catching some exception types
   between libc++ and libstdc++.
 
-**_LIBCPP_INTERNAL_LINKAGE**
-  Mark the affected entity as having internal linkage (i.e. the `static`
-  keyword in C). This is only a best effort: when the `internal_linkage`
-  attribute is not available, we fall back to forcing the function to be
-  inlined, which approximates internal linkage since an externally visible
-  symbol is never generated for that function. This is an internal macro
-  used as an implementation detail by other visibility macros. Never mark
-  a function or a class with this macro directly.
-
-**_LIBCPP_ALWAYS_INLINE**
-  Forces inlining of the function it is applied to. For visibility purposes,
-  this macro is used to make sure that an externally visible symbol is never
-  generated in an object file when the `internal_linkage` attribute is not
-  available. This is an internal macro used by other visibility macros, and
-  it should not be used directly.
-
 Links
 =====
 
 * `[cfe-dev] Visibility in libc++ - 1 <http://lists.llvm.org/pipermail/cfe-dev/2013-July/030610.html>`_
 * `[cfe-dev] Visibility in libc++ - 2 <http://lists.llvm.org/pipermail/cfe-dev/2013-August/031195.html>`_
 * `[libcxx] Visibility fixes for Windows <http://lists.llvm.org/pipermail/cfe-commits/Week-of-Mon-20130805/085461.html>`_

libcxx/docs/ReleaseNotes.rst

 - The ``LIBCXX_ENABLE_DEBUG_MODE_SUPPORT`` CMake configuration is not supported anymore. If you
   were disabling support for the debug mode with that flag, please use ``LIBCXX_ENABLE_BACKWARDS_COMPATIBILITY_DEBUG_MODE_SYMBOLS=OFF``
   instead.
 
 - MinGW DLL builds of libc++ no longer use dllimport in their headers, which
   means that the same set of installed headers works for both DLL and static
   linkage. This means that distributors finally can build both library
   versions with a single CMake invocation.
+
+- The ``LIBCXX_HIDE_FROM_ABI_PER_TU_BY_DEFAULT`` configuration option has been removed. Indeed,
+  the risk of ODR violations from mixing different versions of libc++ in the same program has
+  been mitigated with a different technique that is simpler and does not have the drawbacks of
+  using internal linkage.

libcxx/include/__config

 #ifndef _LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER
 #  pragma GCC system_header
 #endif
 
 #ifdef __cplusplus
 
 #  define _LIBCPP_VERSION 15000
 
+#  define _LIBCPP_CONCAT_IMPL(_X, _Y) _X##_Y
+#  define _LIBCPP_CONCAT(_X, _Y) _LIBCPP_CONCAT_IMPL(_X, _Y)
+
+// Valid C++ identifier that revs with every libc++ version. This can be used to
+// generate identifiers that must be unique for every released libc++ version.
+#  define _LIBCPP_VERSIONED_IDENTIFIER _LIBCPP_CONCAT(v, _LIBCPP_VERSION)
+
 #  if __STDC_HOSTED__ == 0
 #    define _LIBCPP_FREESTANDING
 #  endif
 
 #  ifndef _LIBCPP_STD_VER
 #    if __cplusplus <= 201103L
 #      define _LIBCPP_STD_VER 11
 #    elif __cplusplus <= 201402L

 #    if !defined(_LIBCPP_DISABLE_VISIBILITY_ANNOTATIONS) && __has_attribute(__type_visibility__)
 #      define _LIBCPP_ENUM_VIS __attribute__((__type_visibility__("default")))
 #    else
 #      define _LIBCPP_ENUM_VIS
 #    endif
 
 #  endif // defined(_LIBCPP_OBJECT_FORMAT_COFF)
 
-#  if __has_attribute(internal_linkage)
-#    define _LIBCPP_INTERNAL_LINKAGE __attribute__((internal_linkage))
-#  else
-#    define _LIBCPP_INTERNAL_LINKAGE _LIBCPP_ALWAYS_INLINE
-#  endif
-
 #  if __has_attribute(exclude_from_explicit_instantiation)
 #    define _LIBCPP_EXCLUDE_FROM_EXPLICIT_INSTANTIATION __attribute__((__exclude_from_explicit_instantiation__))
 #  else
 // Try to approximate the effect of exclude_from_explicit_instantiation
 // (which is that entities are not assumed to be provided by explicit
 // template instantiations in the dylib) by always inlining those entities.
 #    define _LIBCPP_EXCLUDE_FROM_EXPLICIT_INSTANTIATION _LIBCPP_ALWAYS_INLINE
 #  endif
 
-#  ifndef _LIBCPP_HIDE_FROM_ABI_PER_TU
-#    ifndef _LIBCPP_HIDE_FROM_ABI_PER_TU_BY_DEFAULT
-#      define _LIBCPP_HIDE_FROM_ABI_PER_TU 0
-#    else
-#      define _LIBCPP_HIDE_FROM_ABI_PER_TU 1
-#    endif
-#  endif
-
-#  ifndef _LIBCPP_HIDE_FROM_ABI
-#    if _LIBCPP_HIDE_FROM_ABI_PER_TU
-#      define _LIBCPP_HIDE_FROM_ABI _LIBCPP_HIDDEN _LIBCPP_INTERNAL_LINKAGE
-#    else
-#      define _LIBCPP_HIDE_FROM_ABI _LIBCPP_HIDDEN _LIBCPP_EXCLUDE_FROM_EXPLICIT_INSTANTIATION
-#    endif
-#  endif
+// This macro marks a symbol as being hidden from libc++'s ABI. This is achieved
+// on two levels:
+// 1. The symbol is given hidden visibility, which ensures that users won't start exporting
+//    symbols from their dynamic library by means of using the libc++ headers. This ensures
+//    that those symbols stay private to the dynamic library in which it is defined.
+//
+// 2. The symbol is given an ABI tag that changes with each version of libc++. This ensures
+//    that no ODR violation can arise from mixing two TUs compiled with different versions
+//    of libc++ where we would have changed the definition of a symbol. If the symbols shared
+//    the same name, the ODR would require that their definitions be token-by-token equivalent,
+//    which basically prevents us from being able to make any change to any function in our
+//    headers. Using this ABI tag ensures that the symbol name is "bumped" artificially at
+//    each release, which lets us change the definition of these symbols at our leisure.
+//    Note that historically, this has been achieved in various ways, including force-inlining
+//    all functions or giving internal linkage to all functions. Both these (previous) solutions
+//    suffer from drawbacks that lead notably to code bloat.
+//
+// Note that we use _LIBCPP_EXCLUDE_FROM_EXPLICIT_INSTANTIATION to ensure that we don't depend
+// on _LIBCPP_HIDE_FROM_ABI methods of classes explicitly instantiated in the dynamic library.
+//
+// TODO: We provide a escape hatch with _LIBCPP_NO_ABI_TAG for folks who want to avoid increasing
+//       the length of symbols with an ABI tag. In practice, we should remove the escape hatch and
+//       use compression mangling instead, see https://github.com/itanium-cxx-abi/cxx-abi/issues/70.
+#  ifndef _LIBCPP_NO_ABI_TAG
+#    define _LIBCPP_HIDE_FROM_ABI                                                                                      \
+      _LIBCPP_HIDDEN _LIBCPP_EXCLUDE_FROM_EXPLICIT_INSTANTIATION                                                       \
+          __attribute__((__abi_tag__(_LIBCPP_TOSTRING(_LIBCPP_VERSIONED_IDENTIFIER))))
+#  else
+#    define _LIBCPP_HIDE_FROM_ABI _LIBCPP_HIDDEN _LIBCPP_EXCLUDE_FROM_EXPLICIT_INSTANTIATION
+#  endif
 
 #  ifdef _LIBCPP_BUILDING_LIBRARY
 #    if _LIBCPP_ABI_VERSION > 1
 #      define _LIBCPP_HIDE_FROM_ABI_AFTER_V1 _LIBCPP_HIDE_FROM_ABI
 #    else
 #      define _LIBCPP_HIDE_FROM_ABI_AFTER_V1
 #    endif
 #  else
 #    define _LIBCPP_HIDE_FROM_ABI_AFTER_V1 _LIBCPP_HIDE_FROM_ABI
 #  endif
 
 // Just so we can migrate to the new macros gradually.
 #  define _LIBCPP_INLINE_VISIBILITY _LIBCPP_HIDE_FROM_ABI
 
 // Inline namespaces are available in Clang/GCC/MSVC regardless of C++ dialect.
 // clang-format off
 #  define _LIBCPP_BEGIN_NAMESPACE_STD namespace std { inline namespace _LIBCPP_ABI_NAMESPACE {
 #  define _LIBCPP_END_NAMESPACE_STD }}
 #  define _VSTD std
 
 _LIBCPP_BEGIN_NAMESPACE_STD _LIBCPP_END_NAMESPACE_STD
 
 #  if _LIBCPP_STD_VER > 14
 #    define _LIBCPP_BEGIN_NAMESPACE_FILESYSTEM                                                                         \
        _LIBCPP_BEGIN_NAMESPACE_STD inline namespace __fs { namespace filesystem {
 #  else
 #    define _LIBCPP_BEGIN_NAMESPACE_FILESYSTEM                                                                         \
       _LIBCPP_BEGIN_NAMESPACE_STD namespace __fs { namespace filesystem {
 #  endif

 #  endif
 
 #  ifndef __cpp_consteval
 #    define _LIBCPP_CONSTEVAL _LIBCPP_CONSTEXPR
 #  else
 #    define _LIBCPP_CONSTEVAL consteval
 #  endif
 
 #  ifdef __GNUC__
 #    define _LIBCPP_NOALIAS __attribute__((__malloc__))
 #  else
 #    define _LIBCPP_NOALIAS
 #  endif
 
 #  if __has_attribute(using_if_exists)
 #    define _LIBCPP_USING_IF_EXISTS __attribute__((using_if_exists))
 #  else
 #    define _LIBCPP_USING_IF_EXISTS
 #  endif
 
 #  ifdef _LIBCPP_CXX03_LANG

libcxx/include/__config_site.in

 //===----------------------------------------------------------------------===//
 //
 // 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
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef _LIBCPP___CONFIG_SITE
 #define _LIBCPP___CONFIG_SITE
 
 #cmakedefine _LIBCPP_ABI_VERSION @_LIBCPP_ABI_VERSION@
 #cmakedefine _LIBCPP_ABI_NAMESPACE @_LIBCPP_ABI_NAMESPACE@
 #cmakedefine _LIBCPP_ABI_FORCE_ITANIUM
 #cmakedefine _LIBCPP_ABI_FORCE_MICROSOFT
-#cmakedefine _LIBCPP_HIDE_FROM_ABI_PER_TU_BY_DEFAULT
 #cmakedefine _LIBCPP_HAS_NO_THREADS
 #cmakedefine _LIBCPP_HAS_NO_MONOTONIC_CLOCK
 #cmakedefine _LIBCPP_HAS_MUSL_LIBC
 #cmakedefine _LIBCPP_HAS_THREAD_API_PTHREAD
 #cmakedefine _LIBCPP_HAS_THREAD_API_EXTERNAL
 #cmakedefine _LIBCPP_HAS_THREAD_API_WIN32
 #cmakedefine _LIBCPP_HAS_THREAD_LIBRARY_EXTERNAL
 #cmakedefine _LIBCPP_DISABLE_VISIBILITY_ANNOTATIONS

libcxx/test/libcxx/strings/basic.string/PR42676.sh.cpp

-//===----------------------------------------------------------------------===//
-//
-// 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
-//
-//===----------------------------------------------------------------------===//
-
-// Regression test for PR42676.
-
-// RUN: %{cxx} %{flags} %s -o %t.exe %{compile_flags} %{link_flags} -D_LIBCPP_HIDE_FROM_ABI_PER_TU
-// RUN: %{run}
-
-#include <memory>
-#include <string>
-
-int main(int, char**) {
-    std::string s1(10u, '-', std::allocator<char>()); (void)s1;
-    std::string s2("hello", std::allocator<char>()); (void)s2;
-    std::string s3("hello"); (void)s3;
-    return 0;
-}

lldb/packages/Python/lldbsuite/test/make/Makefile.rules

ifeq (1,$(USE_LIBSTDCPP))
   # Clang requires an extra flag: -stdlib=libstdc++
   ifneq (,$(findstring clang,$(CC)))
     CXXFLAGS += -stdlib=libstdc++ -DLLDB_USING_LIBSTDCPP
     LDFLAGS += -stdlib=libstdc++
   endif
 endif
 
 ifeq (1,$(USE_LIBCPP))
-  CXXFLAGS += -DLLDB_USING_LIBCPP
+  # TODO: Teach LLDB to handle ABI tags in libc++ namespaces.
+  CXXFLAGS += -DLLDB_USING_LIBCPP -D_LIBCPP_NO_ABI_TAG
   ifeq "$(OS)" "Android"
     # Nothing to do, this is already handled in
     # Android.rules.
   else
     CXXFLAGS += -stdlib=libc++
     LDFLAGS += -stdlib=libc++
   endif
   ifneq (,$(filter $(OS), FreeBSD Linux NetBSD))

llvm/utils/gn/secondary/libcxx/include/BUILD.gn

if (current_toolchain == default_toolchain) {
   write_cmake_config("write_config_site") {
     input = "__config_site.in"
     output = "$libcxx_generated_include_dir/__config_site"
 
     values = [
       "_LIBCPP_ABI_FORCE_ITANIUM=",
       "_LIBCPP_ABI_FORCE_MICROSOFT=",
       "_LIBCPP_EXTRA_SITE_DEFINES=",
-      "_LIBCPP_HIDE_FROM_ABI_PER_TU_BY_DEFAULT=",
       "_LIBCPP_HAS_NO_FILESYSTEM_LIBRARY=",
       "_LIBCPP_HAS_NO_INCOMPLETE_FORMAT=",
       "_LIBCPP_HAS_NO_INCOMPLETE_RANGES=",
       "_LIBCPP_HAS_NO_THREADS=",
       "_LIBCPP_HAS_NO_MONOTONIC_CLOCK=",
       "_LIBCPP_HAS_MUSL_LIBC=",
       "_LIBCPP_HAS_THREAD_API_PTHREAD=",
       "_LIBCPP_HAS_THREAD_API_EXTERNAL=",