[BUG] bionic `<uchar.h>` fails to parse as C++98/C++03 when included directly (`char16_t`/`char32_t` undeclared)

[zbdd6fb8]

Description

Summary

The bionic C header <uchar.h> cannot be parsed as C++98/C++03: it suppresses
its char16_t/char32_t typedefs in all C++ modes (assuming they are built-in
language types), but they only became built-in types in C++11. In pre-C++11
the type names used by c16rtomb, c32rtomb, mbrtoc16, and mbrtoc32
resolve to nothing, causing four parse errors.

This is normally masked because the NDK's default C++ include path resolves
<uchar.h> to the libc++ wrapper (c++/v1/uchar.h), which is intentionally a
no-op before C++11 and never reaches the bionic header. The bionic header is
reached — and fails — in any configuration without the libc++ include path,
including the supported ndk-build configuration APP_STL := none, which adds
-nostdinc++ itself (build/core/stl.mk:22-23).

Environment

• NDK: r30 (30.0.16248370)
• Host: Linux x86_64
• Compiler: the NDK toolchain's clang++
• Target: any (tested aarch64)

Reproducer

ndk-build (APP_STL := none)

// jni/test.cpp
#include <uchar.h>

int foo() {
    return 0;
}

# jni/Android.mk
LOCAL_PATH := $(call my-dir)
include $(CLEAR_VARS)
LOCAL_MODULE := test
LOCAL_SRC_FILES := test.cpp
include $(BUILD_SHARED_LIBRARY)

# jni/Application.mk
APP_ABI := arm64-v8a
APP_STL := none
APP_CPPFLAGS := -std=c++98

$ ndk-build
[arm64-v8a] Compile++      : test <= test.cpp
.../sysroot/usr/include/uchar.h:65:40: error: unknown type name 'char16_t'
.../sysroot/usr/include/uchar.h:74:40: error: unknown type name 'char32_t'
.../sysroot/usr/include/uchar.h:80:17: error: unknown type name 'char16_t'
.../sysroot/usr/include/uchar.h:86:17: error: unknown type name 'char32_t'
4 errors generated.

Command line (equivalent)

$ aarch64-linux-android29-clang++ -std=c++98 -nostdinc++ -c test.cpp

For comparison, the same project builds fine with the default APP_STL
(c++_static) and with APP_STL := system, because in both cases
<uchar.h> resolves to the libc++ wrapper. (Note that APP_STL := system
currently also uses the libc++ headers per the long-standing TODO in
build/core/stl.mk.)

Actual result (r30)

.../sysroot/usr/include/uchar.h:65:40: error: unknown type name 'char16_t'
size_t c16rtomb(char* __buf, char16_t __ch16, mbstate_t* __ps) ...
                                  ^
.../sysroot/usr/include/uchar.h:74:40: error: unknown type name 'char32_t'
.../sysroot/usr/include/uchar.h:80:17: error: unknown type name 'char16_t'
.../sysroot/usr/include/uchar.h:86:17: error: unknown type name 'char32_t'
4 errors generated.

Same failure with -std=c++03. C11 and C++11 and later compile fine.

Expected result

The bionic header should either compile in C++98/C++03, or hide its
declarations in pre-C++11 modes so that it at least remains includable in any
configuration.

Root cause

sysroot/usr/include/uchar.h (from bionic), lines 45-49 in r30:

#if !defined(__cplusplus)
/** The UTF-16 character type. */
typedef __CHAR16_TYPE__ char16_t;
/** The UTF-32 character type. */
typedef __CHAR32_TYPE__ char32_t;
#endif

The typedefs are suppressed in all C++ modes because char16_t/char32_t
are keywords in C++11 and later. But in C++98/C++03 they are not keywords,
and the header no longer typedefs them either, so the type names in the
function declarations below fail to resolve.

By contrast, the libc++ wrapper (c++/v1/uchar.h) treats <uchar.h> as a
C++11-only header: for pre-C++11 it includes nothing (or the frozen C++03
header when _LIBCPP_USE_FROZEN_CXX03_HEADERS is set). That is why default
builds never see the problem — and why the bionic header should not be
reached in that mode, which a __cplusplus guard would enforce.

Suggested fix

Hide the declarations in pre-C++11 modes, consistent with how libc++ treats
the header:

#if !defined(__cplusplus) || __cplusplus >= 201103L
size_t c16rtomb(char* __buf, char16_t __ch16, mbstate_t* __ps) ...;
size_t c32rtomb(char* __buf, char32_t __ch32, mbstate_t* __ps) ...;
size_t mbrtoc16(char16_t* __ch16, const char* __s, size_t __n, mbstate_t* __ps) ...;
size_t mbrtoc32(char32_t* __ch32, const char* __s, size_t __n, mbstate_t* __ps) ...;
#endif

(Alternatively, provide the typedefs for __cplusplus < 201103L so the
declarations parse and the C11 functions remain usable from C++98 — keeping
the keywords untouched for C++11+.)

Notes

The header originates from platform/bionic, so the fix likely needs to land
upstream there and then roll into the NDK sysroot.

I am using a supported NDK

• I have checked and the NDK I'm using is currently supported

Affected versions

r30

[enh-google]

i think i'll go with testing __cpp_unicode_characters to better match what we already did with char8_t:

#if !defined(__cpp_char8_t)
/** The UTF-8 character type. */
typedef unsigned char char8_t;
#endif

#if !defined(__cpp_unicode_characters)
/** The UTF-16 character type. */
typedef __CHAR16_TYPE__ char16_t;
/** The UTF-32 character type. */
typedef __CHAR32_TYPE__ char32_t;
#endif

but otherwise, "yes, will fix in bionic and put this on the list for the next ndk". thanks for the bug report!

[enh-google]

__cpp_unicode_characters change merged in bionic proper... back to the ndk folks...