/***
    Copyright (C) 2023 J Reece Wilson (a/k/a "Reece"). All rights reserved.

    File: ThreadCookie.hpp
    Date: 2023-3-14
    Author: Reece
***/
#pragma once

#if defined(AURORA_IS_MODERNNT_DERIVED) && defined(AURORA_IS_64BIT)
extern "C"
{
    AuUInt32 GetCurrentThreadIDFast();
}
#elif defined(AURORA_IS_LINUX_DERIVED) && defined(AURORA_ARCH_X64)
    unsigned long long __readfsqword(unsigned int) noexcept;
#endif

namespace Aurora::Threading::Primitives
{
#if defined(AURORA_IS_MODERNNT_DERIVED)
    using ThreadCookie_t = AuUInt32; // nt optimization: always use non-pointers (thread ids) as cookies
#else
    using ThreadCookie_t = AuUInt;
#endif

    static auline ThreadCookie_t GetThreadCookie()
    {
    #if defined(AURORA_IS_MODERNNT_DERIVED)
    #if defined(AURORA_ARCH_X86)
        __asm {
            mov eax, fs:[0x18]
            mov eax, [eax + 0x24]
            ret
        }
    #elif defined(AURORA_ARCH_X64)
        return ::GetCurrentThreadIDFast();
    #else
        return GetCurrentThreadId();
    #endif
    #elif defined(AURORA_IS_LINUX_DERIVED) && defined(AURORA_ARCH_X64)
        // _readfsbase_u64 is the modern gcc/intel/msft version of __readgsqword that we need.
        // glibc uses fs[0] = "self" with a note pthreads doesnt use this, fs[1]=dtv counter,fs[2]=self (see: tcbhead_t)
        // musl uses fs[0]  = pSelf, dtv counter, * unused *, * unused * 
        return (ThreadCookie_t)__readfsqword(0);
        //return (ThreadCookie_t)_readfsbase_u64(0);
    #elif defined(AURORA_IS_POSIX_DERIVED)
        // pthread_self is usually a c-friendly a macro in a header that reads from a segment
        // though the result is always a pointer wide
        // this should be faster than the worst case of syscall wrappers of nonstandard get thread id of whatever family of os
        // example: https://elixir.bootlin.com/musl/v1.1.1/source/arch/arm/pthread_arch.h#L14
        //          https://elixir.bootlin.com/musl/v1.1.1/source/arch/x32/pthread_arch.h#L1
        //          powermac segment? https://github.com/phracker/MacOSX-SDKs/blob/041600eda65c6a668f66cb7d56b7d1da3e8bcc93/MacOSX10.3.0.sdk/System/Library/Frameworks/Kernel.framework/Versions/A/Headers/ppc/cpu_capabilities.h#L142
        //          modern mac? https://github.com/apple/darwin-libpthread/blob/2b46cbcc56ba33791296cd9714b2c90dae185ec7/src/pthread.c#L891
        //          x86_32 darwin? movl   %gs:0x0, %edx https://github.com/apple/darwin-libpthread/blob/2b46cbcc56ba33791296cd9714b2c90dae185ec7/src/pthread_asm.s#L162
        // glibc is just slow nested c cancer
        // exactly what youd expect at peak freetard
        // wont be our problem when we start linking against muslc to get away from their crap
        return (ThreadCookie_t)pthread_self();
    #else
        return (ThreadCookie_t)Threads::GetThreadId();
    #endif
    }

}