[Bug tree-optimization/120817] [13/14/15/16 regression] Wrong code when compiled with -O1 -ftree-loop-vectorize for AArch64 target

[tnfchris at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120817

--- Comment #7 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #6)
> A C testcase would be really nice.

Have not been able to reproduce it as C yet, but here's a cut-down C++ version

#include <array>
#include <cstdint>
#include <iostream>

extern int getSafeSize(int inSize, int outSize);

auto copySourceIntoTarget(const std::string& source)
{
    std::array<std::uint8_t, 5> target{};

    const auto loop_end = source.begin() + getSafeSize(source.size(),
target.size());
    auto source_it = source.begin();
    auto target_it = target.begin();
    for (; source_it != loop_end; source_it++, target_it++)
    {
        *target_it = *source_it;
    }

    return target;
}

DSE5 is eliminating the epilogue loop by removing all side-effects from it.

It marks the store

  Deleted dead call: # .MEM_29 = VDEF <.MEM_42>
# USE = anything 
# CLB = anything 
.MASK_STORE (vectp_target.45_76, 8B, max_mask_82, vect__30.43_63);

as dead, even though it's not.

So far my equivalent C version works fine..

void copySourceIntoTarget2 (char * __restrict source, char * __restrict target,
                            int src_begin, int src_size,
                            int target_begin, int target_size)
{
    char* loop_end = source + getSafeSize(src_size, target_size);
    char *source_it = source;
    char *target_it = target;
    for (; source_it != loop_end; source_it++, target_it++)
    {
        *target_it = *source_it;
    }
}

and keeps the epilogue.