Using MSVC, Clang and GCC x86 intrinsics to use x86's SSE4.2 ISA to implement some string manip functions

x86_fast_string.cpp

// x86_fast_string.cpp
#include <nmmintrin.h>
#include <immintrin.h>
#include <cstdint>

#if defined(_MSC_VER)

#define x86_fetch_eflags(mem_dst_32) \
    __asm { \
	__asm pushfd \
	__asm pop eax \
	__asm lea ecx, mem_dst_32 \
	__asm mov [ecx], eax \
    }

#define vec128 __m128i
#define intrin_loadunal_128 _mm_loadu_si128
#define intrin_pcmpistri _mm_cmpistri
#define LOADUNAL128_CAST(expr) ((vec128*)(expr))

#elif defined(__GNUC_MINOR__)

#if defined(__x86_64__)
#define x86_fetch_eflags(mem_dst_32) \
    __asm__ __volatile__( \
            "pushf;" \
            "popq %%rax;" \
            "leaq %0, %%rcx;" \
            "movl %%eax, (%%rcx);" \
            : "=m"(mem_dst_32))
#elif defined(__i386__) || defined(_X86_)
#define x86_fetch_eflags(mem_dst_32) \
    __asm__ __volatile__( \
            "pushf;" \
            "popl %%eax;" \
            "leal %0, %%ecx;" \
            "movl %%eax, (%%ecx);" \
            : "=m"(mem_dst_32))
#endif

#define vec128 __v16qi
#define intrin_loadunal_128 __builtin_ia32_loaddqu
#define intrin_pcmpistri __builtin_ia32_pcmpistri128
#define LOADUNAL128_CAST(expr) ((char*)(expr))

#endif

#if defined(__SSE4_2__) || defined(__AVX__)

//preprocessing-time check for SSE4.2 support
// GCC	OK -- -msse4.2 -- __SSE4_2__ defined
// MSVC  -- force /arch:avx -- __AVX__ defined / _M_IX86_FP should be eq. 2 (?)

bool fast_x86_sse42_streq(const char* str, const char* str1) noexcept {
    constexpr int PCMPISTRI_MODE = _SIDD_CMP_EQUAL_EACH | _SIDD_SBYTE_OPS | _SIDD_NEGATIVE_POLARITY;
    constexpr int END = 16;

    int off = 0;
    int eflags;

    do {
	    vec128 xvec		    = intrin_loadunal_128(LOADUNAL128_CAST(str + off));		//MOVDQU
	    vec128 xvec1		= intrin_loadunal_128(LOADUNAL128_CAST(str1 + off));		//MOVDQU

	    int first_unmatch	= intrin_pcmpistri(xvec, xvec1, PCMPISTRI_MODE);	//PCMPISTRI

	    x86_fetch_eflags(eflags);

	    if (first_unmatch != END)
	        return false;

	    off += 16;
    } while (!(eflags & (1 << 6)) && !(eflags & (1 << 7))); // ZF = 0 and SF = 0

    return true;
}

std::size_t fast_x86_sse42_strlen(const char* nul_str) noexcept {
  constexpr int PCMPISTRI_MODE = _SIDD_CMP_EQUAL_EACH | _SIDD_SBYTE_OPS;

  std::size_t off = 0;
  vec128 zeroed_vec = {0};

  while(true) {
    vec128 str_vec = intrin_loadunal_128(LOADUNAL128_CAST(nul_str + off));

    std::size_t current = intrin_pcmpistri(str_vec, zeroed_vec, PCMPISTRI_MODE);
    if (current < 16) return off + current;

    off += 16;
  }
}

#endif