使用g++ -march=native -g编译AVX架构Intel CPU时触发SIGSEGV

结论 #

这是GCC的代码生成bug，和你的代码无关。核心依据：

• 仅在GCC搭配-march=native/skylake/haswell -g（或无优化但启用AVX相关架构预设）时触发崩溃，Clang编译所有场景均正常
• 崩溃源于GCC生成的vmovdqa指令访问了未按32字节对齐的栈地址，而AVX指令集要求该指令的操作数必须严格32字节对齐
• -O3优化下问题消失，是因为优化逻辑消除了触发错误的栈操作代码

问题细节分析 #

崩溃点的汇编指令vmovdqa %ymm0,-0x40(%rbp)要求rbp-0x40必须是32字节对齐地址，但你的环境中rbp值为0x5fec10，计算后地址未满足对齐要求。

你的测试进一步锁定了问题范围：

• -march=sandybridge -g编译正常：Sandybridge仅支持AVX1，栈对齐要求为16字节，vmovdqa在此模式下会被替换为兼容指令或降低对齐要求
• 手动指定AVX2等独立指令集（-mavx2 -mfma ... -g）编译正常：说明问题并非指令集本身，而是GCC在-march=skylake/haswell这类架构预设配置下的栈对齐处理逻辑错误
• 移除-g但不添加优化时仍崩溃：排除调试信息的直接影响，确认是代码生成阶段的bug

解决建议 #

除了避免使用-march=native，还可以尝试以下方案：

1. 替换架构预设为具体指令集：用-mavx2 -mfma -mbmi -mbmi2 -mmovbe -mlzcnt -mpopcnt -mrdrnd -mf16c替代-march=skylake/haswell，你的测试已验证此方法有效
2. 强制栈对齐编译选项：添加-mstackrealign或-mpreferred-stack-boundary=5（数值5对应32字节对齐），强制GCC生成符合AVX要求的栈对齐代码
3. 升级GCC版本：尝试使用GCC 15.x或更高版本，此类代码生成bug通常会在后续版本中修复
4. 临时启用低级别优化：调试阶段可临时添加-O1，既保留部分调试信息，又可能消除触发崩溃的栈操作逻辑

最小复现代码 #

#include <cstdint>
#include <iostream>

template<typename It>
struct container_view
{
    typedef typename It::reference reference;
    typedef std::remove_reference_t<reference> value_type;
    typedef const reference const_reference;

    constexpr container_view(It it, size_t s) noexcept : it(it), s(s) {}

    constexpr It begin() const noexcept { return it; }
    constexpr It end() const noexcept { return it + size(); }

    constexpr reference operator[](size_t i) noexcept { return *(it + i); }

    /// Size of container.
    constexpr size_t size() const noexcept { return s; }

private:
    It it;
    size_t s;
};


template<typename T, typename S = const size_t>
struct shifted_vector
{
    typedef std::remove_reference_t<T> vector_type;     ///< The type of vector.
    T vec;                                              ///< DenseVector of sequential non-zero elements.
    S offset;                                           ///< Offset of first non-zero element.
};


struct band_matrix_jump_iterator
{
    typedef short value_type;
    typedef short* pointer;
    typedef short& reference;

private:
    size_t b;   // matrix's bandwidth
    size_t v;   // matrix's column
    size_t e;   // matrix's column's element's index.
    pointer p;  // Pointer to current element.

public:

    constexpr band_matrix_jump_iterator(pointer p, size_t b, size_t v, size_t e) noexcept
        : b{b}, v{v}, e{e}, p{p} {}

    constexpr band_matrix_jump_iterator &operator++() noexcept { ++e; ++p; return *this; }
    constexpr band_matrix_jump_iterator &operator+=(ptrdiff_t i) noexcept { e += i; p += i; return *this; }
    constexpr band_matrix_jump_iterator operator+(ptrdiff_t i) const noexcept
        { band_matrix_jump_iterator t(*this); t += i; return t; }
    constexpr reference operator*() const noexcept { return *p; }
    constexpr bool operator==(const band_matrix_jump_iterator &it) const noexcept { return e == it.e; }
};


struct band_matrix
{
    typedef short value_type;
    typedef value_type* pointer;

    constexpr size_t columns() const noexcept { return 10; }

    size_t b;
    pointer elements;


//---------------------------------------------------------------- TYPEDEFS OF ITERATORS AND VECTORS


    typedef band_matrix_jump_iterator column_iterator;
    typedef shifted_vector<container_view<column_iterator>, const uint8_t> column_vector;



//--------------------------------------------------------------------------- CONSTRUCTORS / ASSIGNS

    constexpr band_matrix(size_t b) : b{b}, elements(new value_type[10]) {}
    constexpr ~band_matrix() noexcept { delete[] elements; }


//------------------------------------------ GET ROWS, COLUMNS, DIAGONAL AND CORRESPONDING ITERATORS


    /// Return a beginning iterator to column \c i vector.
    constexpr column_iterator column_begin(size_t i) noexcept
    { return column_iterator(elements + i, 0, i, 0); }


    /// Return a vector view of column \c i.
    constexpr column_vector column(size_t i) noexcept
    { return {typename column_vector::vector_type{column_begin(i), 1}, (uint8_t) i}; }
};



using namespace std;

int main()
{
    band_matrix m(0);
    for (size_t i = 0; i < 10; ++i) m.elements[i] = i;

    // checking columns
    for (size_t i = 0; i < m.columns(); ++i)
    {
        auto v = m.column(i);
        cout << (int) v.offset << ": { ";
        for (auto i : v.vec) cout << i << " ";
        cout << "}\n";
    }
}

编译器版本信息 #

存在问题的GCC版本 #

Using built-in specs.
COLLECT_GCC=g++
COLLECT_LTO_WRAPPER=C:/Program\ Files/mingw64/bin/../libexec/gcc/x86_64-w64-mingw32/14.2.0/lto-wrapper.exe
OFFLOAD_TARGET_NAMES=nvptx-none
Target: x86_64-w64-mingw32
Configured with: ../configure --prefix=/R/winlibs_staging_ucrt64/inst_gcc-14.2.0/share/gcc --build=x86_64-w64-mingw32 --host=x86_64-w64-mingw32 --enable-offload-targets=nvptx-none --with-pkgversion='MinGW-W64 x86_64-ucrt-posix-seh, built by Brecht Sanders, r2' --with-tune=generic --enable-checking=release --enable-threads=posix --disable-sjlj-exceptions --disable-libunwind-exceptions --disable-serial-configure --disable-bootstrap --enable-host-shared --enable-plugin --disable-default-ssp --disable-rpath --disable-libstdcxx-debug --disable-version-specific-runtime-libs --with-stabs --disable-symvers --enable-languages=c,c++,fortran,lto,objc,obj-c++ --disable-gold --disable-nls --disable-stage1-checking --disable-win32-registry --disable-multilib --enable-ld --enable-libquadmath --enable-libada --enable-libssp --enable-libstdcxx --enable-lto --enable-fully-dynamic-string --enable-libgomp --enable-graphite --enable-mingw-wildcard --enable-libstdcxx-time --enable-libstdcxx-pch --with-mpc=/c/Prog/winlibs_staging_ucrt/custombuilt64 --with-mpfr=/c/Prog/winlibs_staging_ucrt/custombuilt64 --with-gmp=/c/Prog/winlibs_staging_ucrt/custombuilt64 --with-isl=/c/Prog/winlibs_staging_ucrt/custombuilt64 --disable-libstdcxx-backtrace --enable-install-libiberty --enable-__cxa_atexit --without-included-gettext --with-diagnostics-color=auto --enable-clocale=generic --with-libiconv --with-system-zlib --with-build-sysroot=/R/winlibs_staging_ucrt64/gcc-14.2.0/build_mingw/mingw-w64 CFLAGS='-D__USE_MINGW_ANSI_STDIO=0 -I/c/Prog/winlibs_staging_ucrt/custombuilt64/include/libdl-win32   -march=nocona -msahf -mtune=generic -O2 -Wno-error=format' CXXFLAGS='-D__USE_MINGW_ANSI_STDIO=0 -Wno-int-conversion  -march=nocona -msahf -mtune=generic -O2' LDFLAGS='-pthread -Wl,--no-insert-timestamp -Wl,--dynamicbase -Wl,--high-entropy-va -Wl,--nxcompat -Wl,--tsaware' LD=/c/Prog/winlibs_staging_ucrt/custombuilt64/share/binutils/bin/ld.exe
Thread model: posix
Supported LTO compression algorithms: zlib zstd
gcc version 14.2.0 (MinGW-W64 x86_64-ucrt-posix-seh, built by Brecht Sanders, r2)

正常工作的Clang版本 #

(built by Brecht Sanders, r2) clang version 19.1.1
Target: x86_64-w64-windows-gnu
Thread model: posix
InstalledDir: C:/Program Files/mingw64/bin

内容的提问来源于stack exchange，提问作者Chameleon