[wasm32 target] Float cast after !! can be wrong when compiled with -x c++

Clang Frontend
1

Hi all

With all llvm/clang versions that support the wasm32 target (8, 9, current build of 10) the following function returns the correct 0.0f when compiled with '-x c' but it wrongfully returns 1.0f when compiled with '-x c++'.

    #ifdef __cplusplus
    extern "C"
    #endif
    float test()
    {
        static unsigned int mask = 2;
        mask |= 4;
        return (float)(!!(mask & 1));
    }

Here is how I built it to a wasm module:
    clang -x c -nostdinc -O3 -target wasm32 -c -o test.c.o test.c
    wasm-ld -no-entry -export=test -o test.c.wasm test.c.o

    clang -x c++ -nostdinc -O3 -target wasm32 -c -o test.cpp.o test.c
    wasm-ld -no-entry -export=test -o test.cpp.wasm test.cpp.o

And here is how I run the test in node.js v12
    node -e "WebAssembly.instantiate(new Uint8Array(require('fs').readFileSync('test.c.wasm'))).then(m => { console.log('c: ' + m.instance.exports.test()); })"
    node -e "WebAssembly.instantiate(new Uint8Array(require('fs').readFileSync('test.cpp.wasm'))).then(m => { console.log('cpp: ' + m.instance.exports.test()); })"

The output I get is
    c: 0
    cpp: 1

And this is the disassembly when running it through the wasm2wat tool from wabt:
    wasm2wat -f test.c.wasm
      (func $test (type 0) (result f32)
        (local i32)
        (i32.store offset=1024
          (i32.const 0)
          (i32.and
            (local.tee 0
              (i32.load offset=1024
                (i32.const 0)))
            (i32.const 5)))
        (f32.convert_i32_s
          (i32.and
            (local.get 0)
            (i32.const 1))))

    wasm2wat -f test.cpp.wasm
      (func $test (type 0) (result f32)
        (local i32)
        (i32.store offset=1024
          (i32.const 0)
          (i32.and
            (local.tee 0
              (i32.load offset=1024
                (i32.const 0)))
            (i32.const 5)))
        (select
          (f32.const 0x1p+0 (;=1;))
          (f32.const 0x0p+0 (;=0;))
          (local.get 0)))

The disassembly of the '-x c' module reads basically what the original C code was.
But it looks like the '-x c++' module returns something like (value_of_mask_at_function_entry ? 1 : 0) which makes no sense to me.
If the code line `mask |= 4` is changed into `mask &= 1`, the function returns 1 the first time it is called, then 0 afterwards which matches that assumption (the '-x c' variant stays correct).

Playing with compilation flag variations like -O0 did not affect the result.

Here's the versions I tested with (all on Win64)
  LLD 10.0.0 / clang version 10.0.0 (trunk) Target: x86_64-pc-windows-msvc
  LLD 8.0.0 / clang version 8.0.0 (tags/RELEASE_800/final) Target: x86_64-pc-windows-msvc
  LLD 9.0.0 / clang version 9.0.0 (tags/RELEASE_900/final) Target: x86_64-pc-windows-msvc

They all produced identical assembly as far as I can tell.

Thanks for checking this out, hopefully this is the right place to report this and greetings from Tokyo,
- Bernhard

2

Hi Bernhard,

Interesting example. `select` instruction returns the operand corresponding to "1" if the input is not zero, and since the value is 6 that is what getting picked. To avoid that the compiler would need to still emit "and" instruction in the C++ mode. I think this is a bug, can you please file it at bugs.llvm.org - assign to me or Thomas Lively.

Note C++ has (unsigned) bool type and in C !! is done on int. IR for C and C++ variants, respectively (this does not explain the output though):

; Function Attrs: nofree norecurse nounwind
define hidden float @test() local_unnamed_addr #0 {
entry:
%0 = load i32, i32* @test.mask, align 4, !tbaa !2
%or = or i32 %0, 4
store i32 %or, i32* @test.mask, align 4, !tbaa !2
%and = and i32 %0, 1
%conv = sitofp i32 %and to float
ret float %conv
}

; Function Attrs: nofree norecurse nounwind
define hidden float @test() local_unnamed_addr #0 {
entry:
%0 = load i32, i32* @_ZZ4testE4mask, align 4, !tbaa !2
%or = or i32 %0, 4
store i32 %or, i32* @_ZZ4testE4mask, align 4, !tbaa !2
%and = and i32 %0, 1
%tobool = icmp ne i32 %and, 0
%conv = uitofp i1 %tobool to float
ret float %conv
}

-Petr