-fshort-wchar is broken with -O2/-Os

[choehwanjin]

Using -fshort-wchar with em++ produces incorrect results that vary depending on the optimization level.

Version of emscripten/emsdk:

emcc (Emscripten gcc/clang-like replacement + linker emulating GNU ld) 5.0.4 (62e22652509fbe7a00609ce48a653d0d66f27ba5)
clang version 23.0.0git (https:/github.com/llvm/llvm-project dcaab6dd99a33edbf98ad1ade0ea0d196b797910)
Target: wasm32-unknown-emscripten
Thread model: posix

Failing command line in full:
I've tested with the following code (
u16.cpp
wchar.cpp
wcslen.cpp
).

Command line:

em++ -fshort-wchar u16.cpp -O0 && env node ./a.out.js
em++ -fshort-wchar u16.cpp -O2 && env node ./a.out.js
em++ -fshort-wchar u16.cpp -Os && env node ./a.out.js

Results:

length: 10 == 10
find: 6 == -1

length: 10 == 10
find: 6 == -1

length: 10 == 56
find: 6 == -1

u16string::size() — fails when -Os is used
u16string::find() — always fails

u16.cpp:

#include <stdio.h>
#include <string>
int main()
{
    std::u16string s = u"0123456789";
    std::u16string::size_type l = s.size();
    printf("length: %d == %zd\n", 10, l);
    std::u16string::size_type p = s.find(u"678");
    printf("find: %d == %zd\n", 6, p);
    return 0;
}

wchar_t behavior is also unstable.
The my_wcslen() function malfunctions when -O2 or -Os is used.

Command line:

make

Results:

em++ -fshort-wchar -c wcslen.cpp -O0
em++ -fshort-wchar -c wchar.cpp -O0
em++ -fshort-wchar -o wchar.js wchar.o wcslen.o
env node ./wchar.js
wchar_t size: 2
len: 10 == 10
em++ -fshort-wchar -c wcslen.cpp -O2
em++ -fshort-wchar -c wchar.cpp -O2
em++ -fshort-wchar -o wchar.js wchar.o wcslen.o
env node ./wchar.js
wchar_t size: 2
len: 10 == 5
em++ -fshort-wchar -c wcslen.cpp -Os
em++ -fshort-wchar -c wchar.cpp -Os
em++ -fshort-wchar -o wchar.js wchar.o wcslen.o
env node ./wchar.js
wchar_t size: 2
len: 10 == 5

wcslen.cpp:

#include <wchar.h>
extern "C" size_t my_wcslen(const wchar_t* s)
{
    const wchar_t* e = s;
    while (*e) {
        ++e;
    }
    return e - s;
}

wchar.cpp:

#include <stdio.h>
#include <wchar.h>
extern "C" size_t my_wcslen(const wchar_t* s);
int
main()
{
    printf("wchar_t size: %zu\n", sizeof(wchar_t));
    wchar_t str[] = L"0123456789";
    size_t l1 = sizeof(str) / sizeof(str[0]) - 1;
    size_t l2 = my_wcslen(str);
    printf("len: %zu == %zu\n", l1, l2);
    return 0;
}

Makefile:

OPTFLAGS =
run: run0 run2 runs
run0: OPTFLAGS = -O0
run0:
        em++ -fshort-wchar -c wcslen.cpp $(OPTFLAGS)
        em++ -fshort-wchar -c wchar.cpp $(OPTFLAGS)
        em++ -fshort-wchar -o wchar.js wchar.o wcslen.o
        env node ./wchar.js
run2: OPTFLAGS = -O2
run2:
        em++ -fshort-wchar -c wcslen.cpp $(OPTFLAGS)
        em++ -fshort-wchar -c wchar.cpp $(OPTFLAGS)
        em++ -fshort-wchar -o wchar.js wchar.o wcslen.o
        env node ./wchar.js
runs: OPTFLAGS = -Os
runs:
        em++ -fshort-wchar -c wcslen.cpp $(OPTFLAGS)
        em++ -fshort-wchar -c wchar.cpp $(OPTFLAGS)
        em++ -fshort-wchar -o wchar.js wchar.o wcslen.o
        env node ./wchar.js

This wcslen bug is reproducible when my_wcslen() is compiled in a separate translation unit.

[sbc100]

The problem is that the emscripten standard library (where wcslen and other functions are defined), is not compiled with the -short-wchar flag. To use this flag for real one would need to add -short-wchar to get_base_cflags in tools/system_libs.py and then rebuild the all the system libraries.

Alternatively we could create a new library directory within the sysroot (e.g. cache/sysroot/lib/short-wchar/libc.a), but I can't imagine there enough users to warrant this extra complexity.

What program are you building the requires this flag? How do you handle this problem when building the host version of this program? Note that this issue also effect the host version gcc/clang. Here is a simplified example that demonstrates the problem and how it is present in the host toolchain too:

$ cat test/c
#include <wchar.h>
#include <stdio.h>

wchar_t foo[1000] = L"hello";

int main() {
  printf("%zu\n", wcslen(foo));
  return 0;
}
$ gcc -fshort-wchar test.c
$ ./a.out
3
$ ./emcc -fshort-wchar test.c
$ node ./a.out.js
3

(Node the program should output 5 if the correct version of wcslen is included).