cross-platform, gcc, clang, msvc, amd64

I would prefer to see something like this in the implementation. The approach was demonstrated in MDEV-20386. For GCC 4, we must work around broken header files.

#include <wmmintrin.h>

#ifdef __GNUC__

__attribute__((target("sse4.2,pclmul")))

#endif

static unsigned crc32c_pclmul(unsigned crc32c, void *data, size_t size)

{

    __m128i a,b,c;

…

    c=_mm_clmulepi64_si128 (a,b,1);

…

}

For GCC 4 (the oldest that we currently support is 4.8.2), some extra hoops are needed to replace the broken #include header:

#include <string.h>

#include <stdint.h>

#if defined __GNUC__ && !defined __clang__ && __GNUC__ < 5

# define _mm_clmulepi64_si128 __builtin_ia32_pclmulqdq128

typedef long long __attribute((vector_size(16))) __m128i;

#else

# include <wmmintrin.h>

#endif

#ifdef __GNUC__

__attribute__((target("sse4.2,pclmul")))

#endif

unsigned crc32c_pclmul(unsigned crc32c, void *data, size_t size)

{

    __m128i a,b,c;

    c=_mm_clmulepi64_si128 (a,b,1);

    return 0;

}

I would find it risky to compile the entire compilation unit with something that enables many SIMD features that may not be available on the least common denominator (SSE2 for AMD64; Intel 80486 or possibly something newer for IA-32). We do not want the compiler to accidentally enable some exotic instructions for any ‘glue’ code that is supposed to run anywhere.

I do not see why anew feature and a workaround for the said feature will help code to be better. It only adds more code, we will need compiler options for clang-cl, and there are already compiler options for ARM etc. So, if anything, I would prefer not to add both attributes and GCC intrinsics for the broken attributes, to reduce complexity, and increase portability.

I can't imagine that compiler will accidentially add pclmul in cpuid test (crc32_pclmul_enabled), which is single function in question in this particular compilation unit