Release mode build issue with gcc-14.2.0

[theComputeKid]

The MacOS release-mode gcc-14 aarch64 CI started failing after this commit: 55b52c6

The log can be seen here:

[ 31%] Building CXX object src/cpu/CMakeFiles/dnnl_cpu.dir/reorder/cpu_reorder_regular_f32_f16.cpp.o
In file included from /Users/runner/work/oneDNN/oneDNN/oneDNN/src/common/dnnl_thread.hpp:24,
                 from /Users/runner/work/oneDNN/oneDNN/oneDNN/src/cpu/reorder/simple_reorder.hpp:25,
                 from /Users/runner/work/oneDNN/oneDNN/oneDNN/src/cpu/reorder/cpu_reorder.hpp:24,
                 from /Users/runner/work/oneDNN/oneDNN/oneDNN/src/cpu/reorder/cpu_reorder_regular_f32_bf16.cpp:18:
In function 'void dnnl::impl::utils::array_copy(T*, const T*, size_t) [with T = long long int]',
    inlined from 'dnnl::impl::dim_t dnnl::impl::cpu::{anonymous}::get_quant_off(const dnnl_dim_t (&)[12], int, int, dnnl::impl::dim_t, dnnl::impl::dim_t, const dnnl::impl::memory_desc_t&)' at /Users/runner/work/oneDNN/oneDNN/oneDNN/src/cpu/reorder/simple_reorder.hpp:198:22,
    inlined from 'dnnl::impl::cpu::simple_reorder_impl<dnnl_f32, dnnl_format_tag_any, dnnl_bf16, dnnl_format_tag_any, true, dnnl::impl::cpu::spec::reference>::execute(const dnnl::impl::cpu::cpu_reorder_pd_t*, const dnnl::impl::exec_ctx_t&)::<lambda(dnnl::impl::dim_t)>' at /Users/runner/work/oneDNN/oneDNN/oneDNN/src/cpu/reorder/simple_reorder.hpp:2617:59:
/Users/runner/work/oneDNN/oneDNN/oneDNN/src/common/utils.hpp:215:16: error: 'void* __builtin_memcpy(void*, const void*, long unsigned int)' specified bound between 18446744056529682432 and 18446744073709551608 exceeds maximum object size 9223372036854775807 [-Werror=stringop-overflow=]
  215 |         dst[i] = src[i];
      |         ~~~~~~~^~~~~~
cc1plus: all warnings being treated as errors
make[2]: *** [src/cpu/CMakeFiles/dnnl_cpu.dir/reorder/cpu_reorder_regular_f32_bf16.cpp.o] Error 1
make[2]: *** Waiting for unfinished jobs....
In file included from /Users/runner/work/oneDNN/oneDNN/oneDNN/src/common/dnnl_thread.hpp:24,
                 from /Users/runner/work/oneDNN/oneDNN/oneDNN/src/cpu/reorder/simple_reorder.hpp:25,
                 from /Users/runner/work/oneDNN/oneDNN/oneDNN/src/cpu/reorder/cpu_reorder.hpp:24,
                 from /Users/runner/work/oneDNN/oneDNN/oneDNN/src/cpu/reorder/cpu_reorder_regular_f32_f16.cpp:17:
In function 'void dnnl::impl::utils::array_copy(T*, const T*, size_t) [with T = long long int]',
    inlined from 'dnnl::impl::dim_t dnnl::impl::cpu::{anonymous}::get_quant_off(const dnnl_dim_t (&)[12], int, int, dnnl::impl::dim_t, dnnl::impl::dim_t, const dnnl::impl::memory_desc_t&)' at /Users/runner/work/oneDNN/oneDNN/oneDNN/src/cpu/reorder/simple_reorder.hpp:198:22,
    inlined from 'dnnl::impl::cpu::simple_reorder_impl<dnnl_f32, dnnl_format_tag_any, dnnl_f16, dnnl_format_tag_any, true, dnnl::impl::cpu::spec::reference>::execute(const dnnl::impl::cpu::cpu_reorder_pd_t*, const dnnl::impl::exec_ctx_t&)::<lambda(dnnl::impl::dim_t)>' at /Users/runner/work/oneDNN/oneDNN/oneDNN/src/cpu/reorder/simple_reorder.hpp:2617:59:
/Users/runner/work/oneDNN/oneDNN/oneDNN/src/common/utils.hpp:215:16: error: 'void* __builtin_memcpy(void*, const void*, long unsigned int)' specified bound between 18446744056529682432 and 18446744073709551608 exceeds maximum object size 9223372036854775807 [-Werror=stringop-overflow=]
  215 |         dst[i] = src[i];
      |         ~~~~~~~^~~~~~
cc1plus: all warnings being treated as errors
make[2]: *** [src/cpu/CMakeFiles/dnnl_cpu.dir/reorder/cpu_reorder_regular_f32_f16.cpp.o] Error 1

We suspect this is a compiler bug and attempts to reproduce it on godbolt did not work, though I was able to reproduce on a local Mac machine. @mgouicem also tried to fix it in #2161, but it did not work. As a solution, we downgraded from gcc-14(.2.0) to gcc-13 in #2160.

Leaving this here for future reference and debugging assistance, when the time comes to upgrade to gcc-14.

[mgouicem]

Hi @theComputeKid. As you were able to reproduce it locally, does the issue also reproduce with the small reproducer I shared on godbolt?
If so it might be worth submitting this issue to gcc bugzilla.

[theComputeKid]

@mgouicem Unfortunately, it did not reproduce with your snippet. But when I next get time, I'll try isolating a repro myself on my machine.

[theComputeKid]

@mgouicem and @vpirogov : just found out that the issue is also present on linux with gcc-14.2 (and not just mac). Are there any changes that can be made to the commit that caused this so that we can compile again?

[dzarukin]

Hi @theComputeKid, we are looking into it. Likely the solution is going to be warning suppression because all the solutions found involve code shuffling which is rather adapting for compiler issue than making the code better itself.

[vpirogov]

+1
It also would be prudent to submit a defect against GCC if it's not a known bug.