10

I am trying to understand this clang-tidy warning: altera-id-dependent-backward-branch that seems to be triggered by this loop.

        for(; first != current; ++first)

The example I have is this code, that looks almost exactly as a perfectly reasonable implementation of std::uninitialized_fill_n.

The static analyzer complains that:

error: backward branch (for loop) is ID-dependent due to variable reference to 'current' and may cause performance degradation [altera-id-dependent-backward-branch,-warnings-as-errors]
                        for(; current != first; ++first) {

uninitialized_fill_n(ForwardIt first, Size n, T const& v) {
    ForwardIt current = first;
    try {
        for(; n > 0; ++current, --n) {
            new (std::addressof(*current)) T(v);
        }
        return current;
    } catch(...) {
        for(; first != current; ++first) {  // clang-tidy error here
            std::addressof(*first)->~T();
        }
        throw;
    }
}

I tried different ways to rewrite the code (for example making the loop backward), to suppress this warning but I couldn't. Is there a standard way to rewrite the fallback (catch) loop in a way that is ok with altera-id-dependent-backward-branch?

I am using clang-tidy 13.0.0.

I also found another instance of the same warning in my code base (this is an implementation of a hash function, by Howard Hinnart):

auto fnv1a(void const* key, std::size_t len, std::size_t h) noexcept {
    auto const *p = static_cast<unsigned char const*>(key);
    unsigned char const* const e = p + len;
    for(; p < e; ++p) {  // warning here
        h = (h ^ *p) * 1099511628211U; // prime
    }
    return h;
}
alfC
  • 14,261
  • 4
  • 67
  • 118
  • 1
    May be interesting: LLVM initial code review for this check: https://reviews.llvm.org/D70094/new/ – yeputons Dec 17 '21 at 18:05
  • I don't get, the examples seem to imply variables that depend on thread id? – alfC Dec 17 '21 at 21:47
  • 1
    Me neither. It seems to me that this is a highly OpenCL-specific diagnostics, I have no idea why it a) made its way into clang-tidy; b) is triggered on your code, I don't see anything that resembles an access to thread id. – yeputons Dec 17 '21 at 22:10
  • I give up, I will disable this diagnosis. Even a "triangular loop" `for(int i = 0; i = N; ++i) for(int j = 0; j != i; ++j)` produces this warning. – alfC Dec 17 '21 at 23:32
  • 1
    I've read some LLVM code and it looks [like a bug](https://github.com/llvm/llvm-project/issues/52790) – yeputons Dec 18 '21 at 09:41

0 Answers0