Suboptimal boolean reduction vectorization

[psiha]

Both Clang and GCC struggle, ARM and x86, but GCC does a better job overall (yet still suboptimal/worse than the handwritten version). If you make the accumulation variable (occurrences) 32 bit at least that helps them use horizontal adds/addv (and it again helps GCC more than Clang).

std::size_t count_occurrences(const std::uint8_t * buf, std::size_t size)
{
    std::size_t occurrences = 0;
    for (auto i = 0; i < size; ++i)
    {
        occurrences += ( buf[ i ] == 'a' );
    }
    return occurrences;
}

https://godbolt.org/z/E13T3MKv7

minbench.zip

[llvmbot]

@llvm/issue-subscribers-backend-arm

Author: Domagoj Šarić (psiha)

Both Clang and GCC struggle, ARM and x86, but GCC does a better job overall (yet still suboptimal/worse than the handwritten version). If you make the accumulation variable (occurrences) 32 bit at least that helps them use horizontal adds/addv (and it again helps GCC more than Clang).

std::size_t count_occurrences(const std::uint8_t * buf, std::size_t size)
{
    std::size_t occurrences = 0;
    for (auto i = 0; i < size; ++i)
    {
        occurrences += ( buf[ i ] == 'a' );
    }
    return occurrences;
}

https://godbolt.org/z/E13T3MKv7

minbench.zip

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Domagoj Šarić (psiha)

Both Clang and GCC struggle, ARM and x86, but GCC does a better job overall (yet still suboptimal/worse than the handwritten version). If you make the accumulation variable (occurrences) 32 bit at least that helps them use horizontal adds/addv (and it again helps GCC more than Clang).

std::size_t count_occurrences(const std::uint8_t * buf, std::size_t size)
{
    std::size_t occurrences = 0;
    for (auto i = 0; i < size; ++i)
    {
        occurrences += ( buf[ i ] == 'a' );
    }
    return occurrences;
}

https://godbolt.org/z/E13T3MKv7

minbench.zip

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Domagoj Šarić (psiha)

Both Clang and GCC struggle, ARM and x86, but GCC does a better job overall (yet still suboptimal/worse than the handwritten version). If you make the accumulation variable (occurrences) 32 bit at least that helps them use horizontal adds/addv (and it again helps GCC more than Clang).

std::size_t count_occurrences(const std::uint8_t * buf, std::size_t size)
{
    std::size_t occurrences = 0;
    for (auto i = 0; i < size; ++i)
    {
        occurrences += ( buf[ i ] == 'a' );
    }
    return occurrences;
}

https://godbolt.org/z/E13T3MKv7

minbench.zip