GCC PCH Breakage Triggered by fmt Scoped `#pragma GCC optimize("Og")`

[bc-lee]

fmt uses a scoped optimization pragma in include/fmt/base.h:

FMT_PRAGMA_GCC(push_options)
#if !defined(__OPTIMIZE__) && !defined(__CUDACC__) && !defined(FMT_MODULE)
FMT_PRAGMA_GCC(optimize("Og"))
#endif
...
FMT_PRAGMA_GCC(pop_options)

This is correct in normal compilation. However, with GCC + PCH, this can leave the compiler in an inconsistent state after PCH restore and break unrelated code.

Example failure:
error: the last argument must be an 8-bit immediate
from _mm_srli_si128(x, 2) in <emmintrin.h>.

The same code compiles fine without PCH.

Related Bugs

I originally reported as GCC bug(RHBZ 2457604): https://bugzilla.redhat.com/show_bug.cgi?id=2457604

Also similar bug report in GCC bugzilla: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105120

Similar bug report suggest that #pragma GCC optimize state is not fully or consistently restored across PCH boundaries.

Workarounds

Current options are:

• avoid using PCH with affected GCC versions
• avoid affected GCC versions

Proposed Workaround in fmt

Please add an opt-out macro to disable the optimization pragma:

#if !defined(FMT_DISABLE_DEBUG_OPTIMIZE_PRAGMA)
FMT_PRAGMA_GCC(push_options)
#if !defined(__OPTIMIZE__) && !defined(__CUDACC__) && !defined(FMT_MODULE)
FMT_PRAGMA_GCC(optimize("Og"))
#endif
#endif // !defined(FMT_DISABLE_DEBUG_OPTIMIZE_PRAGMA)
...
#if !defined(FMT_DISABLE_DEBUG_OPTIMIZE_PRAGMA)
FMT_PRAGMA_GCC(pop_options)
#endif // !defined(FMT_DISABLE_DEBUG_OPTIMIZE_PRAGMA)

Rationale

• This is a QoI/debug optimization, not required for correctness
• Default behavior remains unchanged
• Provides a reliable escape hatch for GCC + PCH users
• Avoids forcing users to drop PCH or fmt usage

[theflashwin]

Hi @bc-lee,

I was not able to reproduce this error locally. I used a Docker container to try and reproduce the error, using the following Dockerfile:

FROM --platform=linux/amd64 fedora:36

RUN dnf install -y gcc-c++ && dnf clean all

WORKDIR /fmt
COPY . .

RUN mkdir -p /repro && \
    printf '#include <fmt/base.h>\n#include <emmintrin.h>\n' > /repro/pch.h && \
    printf '#include <emmintrin.h>\nint shift_right(__m128i x) {\n    __m128i result = _mm_srli_si128(x, 2);\n    return _mm_cvtsi128_si32(result);\n}\n' > /repro/main.cpp

CMD g++ -msse2 -g -O0 -I/fmt/include -x c++-header /repro/pch.h -o /repro/pch.h.gch && \
    echo "PCH compiled" && \
    g++ -msse2 -g -O0 -I/fmt/include -include /repro/pch.h -c /repro/main.cpp -o /dev/null && \
    echo "Compiled OK"

I used Fedora 37 as well (which ships with gcc12) and Fedora 36 (ships with gcc11).

Can you drop your environment specifications and which version of gcc you're experiencing this on?

[bc-lee]

Thanks for the response.

Unfortunately, your reproducer is not the same as the original problem. Your version puts <fmt/base.h> and <emmintrin.h> in the same PCH without an outer #pragma GCC optimize("Og") around the fmt include. I checked that this compiles on both Fedora 36 / GCC 12.2.1 and Fedora 44 / GCC 16.0.1.

The failing case is:

pch.hxx:

#pragma GCC push_options
#pragma GCC optimize("Og")
#include <fmt/base.h>
#pragma GCC pop_options

repro.cpp:

#include <emmintrin.h>

int shift_right(__m128i x) {
  __m128i result = _mm_srli_si128(x, 2);
  return _mm_cvtsi128_si32(result);
}

I verified that this reproduces on both Fedora 36 / GCC 12.2.1 and Fedora 44 / GCC 16.0.1. The PCH builds successfully, but compiling repro.cpp with that PCH fails with:

error: the last argument must be an 8-bit immediate

So the key difference is the active #pragma GCC optimize("Og") state while building the PCH, not merely whether fmt/base.h and <emmintrin.h> appear in the same PCH.

[theflashwin]

Cut a PR here