https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120428
--- Comment #13 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
>
> constexpr std::size_t ProcessChunkSize = BlockSize * OrderSize;
>
> std::array<std::byte, ProcessChunkSize> buffer{};
>
> std::byte* const bytes = reinterpret_cast<std::byte*>(data);
>
> for (std::size_t i = 0; i < TotalSize; i += ProcessChunkSize)
> {
> std::byte* const values = &bytes[i];
>
> for (std::size_t j = 0; j < OrderSize; j++)
> {
> auto* const buffer_chunk = &buffer[j * BlockSize];
> auto* const value_chunk = &values[order[j] * BlockSize];
>
> std::copy(value_chunk, value_chunk + BlockSize, buffer_chunk);
> }
The inner loop is not completely unrolled since std::copy is lowered to
__builtin_memmove instead of __builtin_memcpy(with -mprefer-vector-width=256,
it can be lower to memcpy)
147size: 22-4, last_iteration: 2-2
148 Loop size: 22
149 Estimated size after unrolling: 144-38
150Not unrolling loop 2: contains call and code would grow.
And w/o completely unroll, the redundant copies are not eliminated.
