C/C++ x86-64跨平台CPU指令集检测函数实现问询

Great question! Implementing clean, cross-platform CPU feature detection for x86-64 (supporting both Intel and AMD) is totally achievable by wrapping platform-specific low-level calls behind simple interfaces like hasAVX(). Let’s walk through how to do this, starting with the most involved one—hasAVX()—then cover the others.

Core Background: CPUID and XCR0 #

All x86-64 CPUs expose feature flags via the CPUID instruction. For AVX specifically, we need two critical checks:

1. The CPU itself supports AVX (via a specific bit in the CPUID output)
2. The operating system has enabled AVX support (via the XCR0 register, which controls extended state saving for SIMD registers)

Cross-Platform Implementation #

We’ll start with a generic CPUID wrapper to hide compiler/OS differences, then build each feature check function on top of it.

Step 1: Include Platform-Specific Headers #

First, pull in the necessary intrinsics headers for each OS:

#include <cstdint>

// Platform-specific intrinsics
#ifdef _WIN32
#include <intrin.h>  // For MSVC's CPUID/XGETBV functions
#else
#include <cpuid.h>   // For GCC/Clang's CPUID functions
#endif

Step 2: Generic CPUID Wrapper #

This function abstracts the CPUID call across compilers, supporting both main leaf and subleaf parameters:

void cpuid(uint32_t leaf, uint32_t subleaf, uint32_t& eax, uint32_t& ebx, uint32_t& ecx, uint32_t& edx) {
#ifdef _WIN32
    int regs[4];
    __cpuidex(regs, leaf, subleaf); // MSVC uses __cpuidex for subleaf support
    eax = regs[0];
    ebx = regs[1];
    ecx = regs[2];
    edx = regs[3];
#else
    __cpuid_count(leaf, subleaf, eax, ebx, ecx, edx); // GCC/Clang's subleaf-aware CPUID
#endif
}

Step 3: Implement hasAVX() #

As noted, AVX requires both CPU and OS support:

bool hasAVX() {
    uint32_t eax, ebx, ecx, edx;
    
    // Check if CPU supports AVX (ECX bit 28 in CPUID leaf 1)
    cpuid(1, 0, eax, ebx, ecx, edx);
    if (!(ecx & (1 << 28))) {
        return false;
    }

    // Check if OS has enabled AVX (XCR0 bits 1 and 2 must be set for XMM/YMM state saving)
#ifdef _WIN32
    uint64_t xcr0 = _xgetbv(0);
#else
    uint64_t xcr0 = __xgetbv(0);
#endif
    return (xcr0 & 0x6) == 0x6;
}

Step 4: Implement Other Feature Checks #

The pattern is similar for SSE3, SSE4.1/4.2, and AVX2—just check the correct CPUID bits (and XCR0 for AVX2):

bool hasSSE3() {
    uint32_t eax, ebx, ecx, edx;
    cpuid(1, 0, eax, ebx, ecx, edx);
    return (ecx & (1 << 0)) != 0;  // ECX bit 0 = SSE3 support
}

bool hasSSE4_1() {
    uint32_t eax, ebx, ecx, edx;
    cpuid(1, 0, eax, ebx, ecx, edx);
    return (ecx & (1 << 19)) != 0; // ECX bit 19 = SSE4.1 support
}

bool hasSSE4_2() {
    uint32_t eax, ebx, ecx, edx;
    cpuid(1, 0, eax, ebx, ecx, edx);
    return (ecx & (1 << 20)) != 0; // ECX bit 20 = SSE4.2 support
}

bool hasAVX2() {
    uint32_t eax, ebx, ecx, edx;
    
    // Check CPU support: AVX2 bit (EBX bit 5) in CPUID leaf 7, subleaf 0
    cpuid(7, 0, eax, ebx, ecx, edx);
    if (!(ebx & (1 << 5))) {
        return false;
    }

    // Same OS support check as AVX
#ifdef _WIN32
    uint64_t xcr0 = _xgetbv(0);
#else
    uint64_t xcr0 = __xgetbv(0);
#endif
    return (xcr0 & 0x6) == 0x6;
}

Usage Example: Runtime Code Path Dispatch #

Once you have these functions, you can easily switch between optimized code paths at runtime:

void processData(const float* input, float* output, size_t size) {
    if (hasAVX2()) {
        processDataAVX2(input, output, size);
    } else if (hasAVX()) {
        processDataAVX(input, output, size);
    } else if (hasSSE4_2()) {
        processDataSSE42(input, output, size);
    } else {
        processDataFallback(input, output, size);
    }
}

Key Notes #

• Compiler Compatibility: This code works with MSVC, GCC, and Clang—all major compilers for Windows, macOS, and Linux.
• AMD/Intel Compatibility: The CPUID feature bits used are part of the x86-64 standard, so they work seamlessly for both AMD and Intel processors.
• OS Support: Skipping the XCR0 check can lead to crashes on systems where the OS hasn’t enabled AVX/AVX2 state saving—don’t omit this step!
• Legacy CPUs: For very old CPUs that don’t support certain CPUID leaves (like leaf 7 for AVX2), the cpuid call will return 0 in the relevant registers, so the feature check will correctly return false.

内容的提问来源于stack exchange，提问作者Royi