[libc++] Only include_next C library headers when they exist

[ldionne]

Summary

Some platforms don't provide all C library headers. In practice, libc++
only requires a few C library headers to exist, and only a few functions
on those headers. Missing functions that libc++ doesn't need for its own
implementation are handled properly by the using_if_exists attribute,
however a missing header is currently a hard error when we try to
do #include_next.

This patch should make libc++ more flexible on platforms that do not
provide C headers that libc++ doesn't actually require for its own
implementation. The only downside is that it may move some errors from
the #include_next point to later in the compilation if we actually try
to use something that isn't provided, which could be somewhat confusing.
However, these errors should be caught by folks trying to port libc++
over to a new platform (when running the libc++ test suite), not by end
users.

NOTE: This is a reapplicaton of 226409, which was reverted in 674729813 because it broke the build. The issue has now been fixed with https://reviews.llvm.org/D138062.

[philnik]

In essence LGTM. As another PR, could we maybe add a minimal set of C headers that are required to build and test libc++ with LIBCXX_ENABLE_LOCALIZATION=Off, LIBCXX_ENABLE_UNICODE=Off, etc? That would allow us to track what we require the C implementation to provide and we can actually test against a minimal C library. If you think that's a good idea I'd be happy to work on that.

libcxx/include/complex.h

28–31

[ldionne]

In D136683#3882795, @philnik wrote:

In essence LGTM. As another PR, could we maybe add a minimal set of C headers that are required to build and test libc++ with LIBCXX_ENABLE_LOCALIZATION=Off, LIBCXX_ENABLE_UNICODE=Off, etc? That would allow us to track what we require the C implementation to provide and we can actually test against a minimal C library. If you think that's a good idea I'd be happy to work on that.

Yes, I think that would make sense, but I think it's also a non-negligible amount of work. You're thinking about making that part of the test suite, right?

[ldionne]

Address review comments and rebase.

[thakis]

It it possible that this caused this check-clang failure: http://45.33.8.238/macm1/48659/step_7.txt ?

[thakis]

In D136683#3927871, @thakis wrote:

It it possible that this caused this check-clang failure: http://45.33.8.238/macm1/48659/step_7.txt ?

git bisect says it's indeed this commit.

[thakis]

Reverted in 674729813e3be02af7bda02cfe577398763f58b9 for now.

[philnik]

In D136683#3927771, @ldionne wrote:

In D136683#3882795, @philnik wrote:

In essence LGTM. As another PR, could we maybe add a minimal set of C headers that are required to build and test libc++ with LIBCXX_ENABLE_LOCALIZATION=Off, LIBCXX_ENABLE_UNICODE=Off, etc? That would allow us to track what we require the C implementation to provide and we can actually test against a minimal C library. If you think that's a good idea I'd be happy to work on that.

Yes, I think that would make sense, but I think it's also a non-negligible amount of work. You're thinking about making that part of the test suite, right?

Yes, that would be the idea. My idea was to write custom headers and use the actual implementation from llvm-libc. I think that should be quite manageable.

[ldionne]

In D136683#3927871, @thakis wrote:

It it possible that this caused this check-clang failure: http://45.33.8.238/macm1/48659/step_7.txt ?

How does that bot configure the build with CMake?

[ldionne]

In D136683#3928001, @ldionne wrote:

In D136683#3927871, @thakis wrote:

It it possible that this caused this check-clang failure: http://45.33.8.238/macm1/48659/step_7.txt ?

How does that bot configure the build with CMake?

@thakis To make sure you see this.

[thakis]

In D136683#3928074, @ldionne wrote:

In D136683#3928001, @ldionne wrote:

In D136683#3927871, @thakis wrote:

It it possible that this caused this check-clang failure: http://45.33.8.238/macm1/48659/step_7.txt ?

How does that bot configure the build with CMake?

@thakis To make sure you see this.

It doesn't. But from the error message, it looks like it shouldn't matter (?)

My reading of the error output is that this test tests the error message that occurs if the c++ headers aren't present. Previously we'd say "file not found", but now we say "reference to unresolved using declaration" due to not including some C header.

Maybe I misunderstood what's happening though.

[ldionne]

In D136683#3928129, @thakis wrote:

It doesn't. But from the error message, it looks like it shouldn't matter (?)

My reading of the error output is that this test tests the error message that occurs if the c++ headers aren't present. Previously we'd say "file not found", but now we say "reference to unresolved using declaration" due to not including some C header.

Maybe I misunderstood what's happening though.

Well that build picked up the latest libc++ headers, so I guess it's at least doing a bootstrapping build and then using check-clang? I don't even know if Clang will use the just-built libc++ for testing when doing a bootstrapping build.

[thakis]

Well that build picked up the latest libc++ headers, so I guess it's at least doing a bootstrapping build and then using check-clang? I don't even know if Clang will use the just-built libc++ for testing when doing a bootstrapping build.

That bot "installs" libc++ headers to buildir/include/c++/v1 and puts clang in builddir/bin/clang, and clang looks for includes at ../../includes relative to it. But I think that's the normal directory layout when doing a make install with the cmake build too (?)

[thakis]

Here's how to repro this failure with the cmake build (can probably repro with a shorter cmake invocation too):

cd my-build-dir
/Applications/CMake.app/Contents/bin/cmake -GNinja -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=ON -DLLVM_ENABLE_PROJECTS='clang;lld' -DLLVM_ENABLE_RUNTIMES='compiler-rt;libcxx' -DLLVM_APPEND_VC_REV=NO -DLLVM_TARGETS_TO_BUILD='X86;ARM;AArch64' ../llvm-project/llvm

ninja clang
ninja cxx
ninja FileCheck count not     

bin/llvm-lit ../llvm-project/clang/test/Driver/nostdincxx.cpp -vv

[ldionne]

@thakis https://reviews.llvm.org/D138062 was the problem. The test was written incorrectly, basically.

[ldionne]

Rebase and reopen.

[ldionne]

Rebase

[ldionne]

@thakis I'm going to land this tomorrow if the Clang change sticks.

[thakis]

I just noticed that this leads to atrocious diagnostics:

% cat hello.cc
#include <stdio.h>
int main() { printf("hello\n"); }

% out/gn/bin/clang -c hello.cc      
hello.cc:2:14: error: use of undeclared identifier 'printf'

???

This works:

% out/gn/bin/clang -c hello.cc -isysroot $(xcrun -show-sdk-path)

The reason is that stdio.h picks up libc++'s stdio.h, which then only include_next's the actual stdio.h if it exists.

This seems suboptimal (?)

[ldionne]

In D136683#4013790, @thakis wrote:

I just noticed that this leads to atrocious diagnostics:

% cat hello.cc
#include <stdio.h>
int main() { printf("hello\n"); }

% out/gn/bin/clang -c hello.cc      
hello.cc:2:14: error: use of undeclared identifier 'printf'

???

This works:

% out/gn/bin/clang -c hello.cc -isysroot $(xcrun -show-sdk-path)

The reason is that stdio.h picks up libc++'s stdio.h, which then only include_next's the actual stdio.h if it exists.

This seems suboptimal (?)

Yes, I've noticed that before, but I don't see a solution to this other than getting rid of libc++'s stdio.h header entirely (and same for other C compatibility headers). This is a tradeoff between diagnostics quality and libc++'s ability to work on top of freestanding implementations without much work.

libcxx/include/complex.h

 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
 #ifdef __cplusplus
-
-#include <ccomplex>
-
-#else  // __cplusplus
-
-#include_next <complex.h>
-
-#endif // __cplusplus
+#  include <ccomplex>
+#elif __has_include_next(<complex.h>)
+#  include_next <complex.h>
+#endif
 
 #endif // _LIBCPP_COMPLEX_H

libcxx/include/ctype.h

 */
 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
-#include_next <ctype.h>
+#if __has_include_next(<ctype.h>)
+#  include_next <ctype.h>
+#endif
 
 #ifdef __cplusplus
 
 #undef isalnum
 #undef isalpha
 #undef isblank
 #undef iscntrl
 #undef isdigit

libcxx/include/errno.h

 */
 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
-#include_next <errno.h>
+#if __has_include_next(<errno.h>)
+#  include_next <errno.h>
+#endif
 
 #ifdef __cplusplus
 
 #if !defined(EOWNERDEAD) || !defined(ENOTRECOVERABLE)
 
 #ifdef ELAST
 
 static const int __elast1 = ELAST+1;

libcxx/include/fenv.h

 */
 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
-#include_next <fenv.h>
+#if __has_include_next(<fenv.h>)
+#  include_next <fenv.h>
+#endif
 
 #ifdef __cplusplus
 
 extern "C++" {
 
 #ifdef feclearexcept
 #undef feclearexcept
 #endif

libcxx/include/float.h

 */
 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
-#include_next <float.h>
+#if __has_include_next(<float.h>)
+#  include_next <float.h>
+#endif
 
 #ifdef __cplusplus
 
 #ifndef FLT_EVAL_METHOD
 #define FLT_EVAL_METHOD __FLT_EVAL_METHOD__
 #endif
 
 #ifndef DECIMAL_DIG
 #define DECIMAL_DIG __DECIMAL_DIG__
 #endif
 
 #endif // __cplusplus
 
 #endif // _LIBCPP_FLOAT_H

libcxx/include/inttypes.h

 
 /* C99 stdlib (e.g. glibc < 2.18) does not provide format macros needed
    for C++11 unless __STDC_FORMAT_MACROS is defined
 */
 #if defined(__cplusplus) && !defined(__STDC_FORMAT_MACROS)
 #   define __STDC_FORMAT_MACROS
 #endif
 
-#include_next <inttypes.h>
+#if __has_include_next(<inttypes.h>)
+#  include_next <inttypes.h>
+#endif
 
 #ifdef __cplusplus
 
 #include <stdint.h>
 
 #undef imaxabs
 #undef imaxdiv
 
 #endif // __cplusplus
 
 #endif // _LIBCPP_INTTYPES_H

libcxx/include/limits.h

 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
 #ifndef __GNUC__
-#include_next <limits.h>
+
+#  if __has_include_next(<limits.h>)
+#    include_next <limits.h>
+#  endif
+
 #else
 // GCC header limits.h recursively includes itself through another header called
 // syslimits.h for some reason. This setup breaks down if we directly
 // #include_next GCC's limits.h (reasons not entirely clear to me). Therefore,
 // we manually re-create the necessary include sequence below:
 
 // Get the system limits.h defines (force recurse into the next level)
 #define _GCC_LIMITS_H_

libcxx/include/locale.h

 #if defined(_LIBCPP_HAS_NO_LOCALIZATION)
 #   error "<locale.h> is not supported since libc++ has been configured without support for localization."
 #endif
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
-#include_next <locale.h>
+#if __has_include_next(<locale.h>)
+#  include_next <locale.h>
+#endif
 
 #endif // _LIBCPP_LOCALE_H

libcxx/include/math.h

 */
 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
-#include_next <math.h>
+#  if __has_include_next(<math.h>)
+#    include_next <math.h>
+#  endif
 
 #ifdef __cplusplus
 
 // We support including .h headers inside 'extern "C"' contexts, so switch
 // back to C++ linkage before including these C++ headers.
 extern "C++" {
 
 #include <__type_traits/promote.h>

libcxx/include/setjmp.h

 */
 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
-#include_next <setjmp.h>
+#if __has_include_next(<setjmp.h>)
+#  include_next <setjmp.h>
+#endif
 
 #ifdef __cplusplus
 
 #ifndef setjmp
 #define setjmp(env) setjmp(env)
 #endif
 
 #endif // __cplusplus
 
 #endif // _LIBCPP_SETJMP_H

libcxx/include/stdbool.h

 */
 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
-#include_next <stdbool.h>
+#if __has_include_next(<stdbool.h>)
+#  include_next <stdbool.h>
+#endif
 
 #ifdef __cplusplus
 #undef bool
 #undef true
 #undef false
 #undef __bool_true_false_are_defined
 #define __bool_true_false_are_defined 1
 #endif
 
 #endif // _LIBCPP_STDBOOL_H

libcxx/include/stddef.h

 */
 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
-#include_next <stddef.h>
+#  if __has_include_next(<stddef.h>)
+#    include_next <stddef.h>
+#  endif
 
 #ifdef __cplusplus
     typedef decltype(nullptr) nullptr_t;
 #endif
 
 #endif // _LIBCPP_STDDEF_H

libcxx/include/stdint.h

 */
 #if defined(__cplusplus) && !defined(__STDC_LIMIT_MACROS)
 #   define __STDC_LIMIT_MACROS
 #endif
 #if defined(__cplusplus) && !defined(__STDC_CONSTANT_MACROS)
 #   define __STDC_CONSTANT_MACROS
 #endif
 
-#include_next <stdint.h>
+#if __has_include_next(<stdint.h>)
+#  include_next <stdint.h>
+#endif
 
 #endif // _LIBCPP_STDINT_H

libcxx/include/stdio.h

 */
 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
-#include_next <stdio.h>
+#  if __has_include_next(<stdio.h>)
+#    include_next <stdio.h>
+#  endif
 
 #ifdef __cplusplus
 
 #undef getc
 #undef putc
 #undef clearerr
 #undef feof
 #undef ferror
 
 #endif
 
 #endif // _LIBCPP_STDIO_H

libcxx/include/stdlib.h

 */
 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
-#include_next <stdlib.h>
+#  if __has_include_next(<stdlib.h>)
+#    include_next <stdlib.h>
+#  endif
 
 #ifdef __cplusplus
 extern "C++" {
 // abs
 
 #ifdef abs
 # undef abs
 #endif

libcxx/include/string.h

 */
 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
-#include_next <string.h>
+#if __has_include_next(<string.h>)
+#  include_next <string.h>
+#endif
 
 // MSVCRT, GNU libc and its derivates may already have the correct prototype in
 // <string.h>. This macro can be defined by users if their C library provides
 // the right signature.
 #if defined(__CORRECT_ISO_CPP_STRING_H_PROTO) || defined(_LIBCPP_MSVCRT) || \
     defined(__sun__) || defined(_STRING_H_CPLUSPLUS_98_CONFORMANCE_)
 #define _LIBCPP_STRING_H_HAS_CONST_OVERLOADS
 #endif

libcxx/include/tgmath.h

 
 #include <__config>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
 #ifdef __cplusplus
-
-#include <ctgmath>
-
-#else  // __cplusplus
-
-#include_next <tgmath.h>
-
-#endif // __cplusplus
+#  include <ctgmath>
+#else
+#  if __has_include_next(<tgmath.h>)
+#    include_next <tgmath.h>
+#  endif
+#endif
 
 #endif // _LIBCPP_TGMATH_H

libcxx/include/wchar.h

 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
 #ifdef __cplusplus
 #define __CORRECT_ISO_CPP_WCHAR_H_PROTO
 #endif
 
-#include_next <wchar.h>
+#  if __has_include_next(<wchar.h>)
+#    include_next <wchar.h>
+#  endif
 
 // Determine whether we have const-correct overloads for wcschr and friends.
 #if defined(_WCHAR_H_CPLUSPLUS_98_CONFORMANCE_)
 #  define _LIBCPP_WCHAR_H_HAS_CONST_OVERLOADS 1
 #elif defined(__GLIBC_PREREQ)
 #  if __GLIBC_PREREQ(2, 10)
 #    define _LIBCPP_WCHAR_H_HAS_CONST_OVERLOADS 1
 #  endif