// This file is part of AsmJit project <https://asmjit.com>
//
// See <asmjit/core.h> or LICENSE.md for license and copyright information
// SPDX-License-Identifier: Zlib

#include <asmjit/core/api-build_p.h>
#include <asmjit/core/cpuinfo.h>
#include <asmjit/support/support.h>

#include <atomic>

// Required by `__cpuidex()` and `_xgetbv()`.
#if ASMJIT_ARCH_X86
  #if defined(_MSC_VER)
    #include <intrin.h>
  #endif
#endif // ASMJIT_ARCH_X86

#if ASMJIT_ARCH_ARM
  // Required by various utilities that are required by features detection.
  #if !defined(_WIN32)
    #include <errno.h>
    #include <sys/utsname.h>
  #endif

  //! Required to detect CPU and features on Apple platforms.
  #if defined(__APPLE__)
    #include <mach/machine.h>
    #include <sys/types.h>
    #include <sys/sysctl.h>
  #endif

  #if (defined(__linux__) || defined(__FreeBSD__))
    // Required by `getauxval()` on Linux and FreeBSD.
    #include <sys/auxv.h>
    #define ASMJIT_ARM_DETECT_VIA_HWCAPS
  #endif

  #if ASMJIT_ARCH_ARM >= 64 && defined(__GNUC__) && defined(__linux__) && 0
    // This feature is disabled at the moment - it works, but it seems linux supports ARM features
    // via HWCAPS pretty well and the most recent features need to access more registers that were
    // not originally accessible, which would break on some systems.
    #define ASMJIT_ARM_DETECT_VIA_CPUID
  #endif

  #if ASMJIT_ARCH_ARM >= 64 && defined(__OpenBSD__)
    #include <machine/cpu.h>
    #include <sys/sysctl.h>
  #endif

  #if ASMJIT_ARCH_ARM >= 64 && defined(__NetBSD__)
    #include <sys/sysctl.h>
  #endif

#endif // ASMJIT_ARCH_ARM

#if !defined(_WIN32) && (ASMJIT_ARCH_X86 || ASMJIT_ARCH_ARM)
  #include <unistd.h>
#endif

ASMJIT_BEGIN_NAMESPACE

// CpuInfo - Detect - Compatibility
// ================================

// CPU features detection is a minefield on non-X86 platforms. The following list describes which
// operating systems and architectures are supported and the status of the implementation:
//
//   * X86|X86_64:
//     - All OSes supported
//     - Detection is based on using a CPUID instruction, which is a user-space instruction, so there
//       is no need to use any OS specific APIs or syscalls to detect all features provided by the CPU.
//
//   * ARM32:
//     - Linux   - HWCAPS based detection.
//     - FreeBSD - HWCAPS based detection (shared with Linux code).
//     - NetBSD  - NOT IMPLEMENTED!
//     - OpenBSD - NOT IMPLEMENTED!
//     - Apple   - sysctlbyname() based detection (this architecture is deprecated on Apple HW).
//     - Windows - IsProcessorFeaturePresent() based detection (only detects a subset of features).
//     - Others  - NOT IMPLEMENTED!
//
//   * ARM64:
//     - Linux   - HWCAPS and CPUID based detection.
//     - FreeBSD - HWCAPS and CPUID based detection (shared with Linux code).
//     - NetBSD  - CPUID based detection (reading CPUID via sysctl's cpu0 info)
//     - OpenBSD - CPUID based detection (reading CPUID via sysctl's CTL_MACHDEP).
//     - Apple   - sysctlbyname() based detection with FamilyId matrix (record for each family id).
//     - Windows - IsProcessorFeaturePresent() based detection (only detects a subset of features).
//     - Others  - NOT IMPLEMENTED!
//
//   * Others
//     - NOT IMPLEMENTED!

// CpuInfo - Detect - HW-Thread Count
// ==================================

#if defined(_WIN32)
static inline uint32_t detect_hw_thread_count() noexcept {
  SYSTEM_INFO info;
  ::GetSystemInfo(&info);
  return info.dwNumberOfProcessors;
}
#elif defined(_SC_NPROCESSORS_ONLN)
static inline uint32_t detect_hw_thread_count() noexcept {
  long res = ::sysconf(_SC_NPROCESSORS_ONLN);
  return res <= 0 ? uint32_t(1) : uint32_t(res);
}
#else
static inline uint32_t detect_hw_thread_count() noexcept {
  return 1;
}
#endif

// CpuInfo - Detect - X86
// ======================

// X86 and X86_64 detection is based on CPUID.

#if ASMJIT_ARCH_X86

namespace x86 {

using Ext = CpuFeatures::X86;

struct cpuid_t { uint32_t eax, ebx, ecx, edx; };
struct xgetbv_t { uint32_t eax, edx; };

// Executes `cpuid` instruction.
static inline void cpuid_query(cpuid_t* out, uint32_t in_eax, uint32_t in_ecx = 0) noexcept {
#if defined(_MSC_VER)
  __cpuidex(reinterpret_cast<int*>(out), in_eax, in_ecx);
#elif defined(__GNUC__) && ASMJIT_ARCH_X86 == 32
  __asm__ __volatile__(
    "mov %%ebx, %%edi\n"
    "cpuid\n"
    "xchg %%edi, %%ebx\n" : "=a"(out->eax), "=D"(out->ebx), "=c"(out->ecx), "=d"(out->edx) : "a"(in_eax), "c"(in_ecx));
#elif defined(__GNUC__) && ASMJIT_ARCH_X86 == 64
  __asm__ __volatile__(
    "mov %%rbx, %%rdi\n"
    "cpuid\n"
    "xchg %%rdi, %%rbx\n" : "=a"(out->eax), "=D"(out->ebx), "=c"(out->ecx), "=d"(out->edx) : "a"(in_eax), "c"(in_ecx));
#else
  #error "[asmjit] x86::cpuid_query() - Unsupported compiler."
#endif
}

// Executes 'xgetbv' instruction.
static inline void xgetbv_query(xgetbv_t* out, uint32_t in_ecx) noexcept {
#if defined(_MSC_VER)
  uint64_t value = _xgetbv(in_ecx);
  out->eax = uint32_t(value & 0xFFFFFFFFu);
  out->edx = uint32_t(value >> 32);
#elif defined(__GNUC__)
  uint32_t out_eax;
  uint32_t out_edx;

  // Replaced, because the world is not perfect:
  //   __asm__ __volatile__("xgetbv" : "=a"(out_eax), "=d"(out_edx) : "c"(in_ecx));
  __asm__ __volatile__(".byte 0x0F, 0x01, 0xD0" : "=a"(out_eax), "=d"(out_edx) : "c"(in_ecx));

  out->eax = out_eax;
  out->edx = out_edx;
#else
  out->eax = 0;
  out->edx = 0;
#endif
}

// Map a 12-byte vendor string returned by `cpuid` into a `CpuInfo::Vendor` ID.
static inline void simplify_cpu_vendor(CpuInfo& cpu, uint32_t d0, uint32_t d1, uint32_t d2) noexcept {
  struct Vendor {
    char normalized[8];
    union { char text[12]; uint32_t d[3]; };
  };

  static const Vendor table[] = {
    { { 'A', 'M', 'D'                     }, {{ 'A', 'u', 't', 'h', 'e', 'n', 't', 'i', 'c', 'A', 'M', 'D' }} },
    { { 'I', 'N', 'T', 'E', 'L'           }, {{ 'G', 'e', 'n', 'u', 'i', 'n', 'e', 'I', 'n', 't', 'e', 'l' }} },
    { { 'V', 'I', 'A'                     }, {{ 'C', 'e', 'n', 't', 'a', 'u', 'r', 'H', 'a', 'u', 'l', 's' }} },
    { { 'V', 'I', 'A'                     }, {{ 'V', 'I', 'A',  0 , 'V', 'I', 'A',  0 , 'V', 'I', 'A',  0  }} },
    { { 'U', 'N', 'K', 'N', 'O', 'W', 'N' }, {{ 0                                                          }} }
  };

  uint32_t i;
  for (i = 0; i < ASMJIT_ARRAY_SIZE(table) - 1; i++) {
    if (table[i].d[0] == d0 && table[i].d[1] == d1 && table[i].d[2] == d2) {
      break;
    }
  }
  memcpy(cpu._vendor.str, table[i].normalized, 8);
}

static ASMJIT_FAVOR_SIZE void simplify_cpu_brand(char* s) noexcept {
  char* d = s;

  char c = s[0];
  char prev = 0;

  // Used to always clear the current character to ensure that the result
  // doesn't contain garbage after a new null terminator is placed at the end.
  s[0] = '\0';

  for (;;) {
    if (!c) {
      break;
    }

    if (!(c == ' ' && (prev == '@' || s[1] == ' ' || s[1] == '@' || s[1] == '\0'))) {
      *d++ = c;
      prev = c;
    }

    c = *++s;
    s[0] = '\0';
  }

  d[0] = '\0';
}

static ASMJIT_FAVOR_SIZE void detect_x86_cpu(CpuInfo& cpu) noexcept {
  using Support::bit_test;

  cpuid_t regs;
  xgetbv_t xcr0 { 0, 0 };
  CpuFeatures::X86& features = cpu.features().x86();

  cpu._was_detected = true;
  cpu._max_logical_processors = 1;

  // We are gonna execute CPUID, which was introduced by I486, so it's the requirement.
  features.add(Ext::kI486);

  // CPUID EAX=0x00 (Basic CPUID Information)
  // ----------------------------------------

  // Get vendor string/id.
  cpuid_query(&regs, 0x0);

  uint32_t max_id = regs.eax;
  uint32_t max_sub_leaf_id_0x7 = 0;

  simplify_cpu_vendor(cpu, regs.ebx, regs.edx, regs.ecx);

  // CPUID EAX=0x01 (Basic CPUID Information)
  // ----------------------------------------

  if (max_id >= 0x01u) {
    // Get feature flags in ECX/EDX and family/model in EAX.
    cpuid_query(&regs, 0x1);

    // Fill family and model fields.
    uint32_t model_id  = (regs.eax >> 4) & 0x0F;
    uint32_t family_id = (regs.eax >> 8) & 0x0F;

    // Use extended family and model fields.
    if (family_id == 0x06u || family_id == 0x0Fu) {
      model_id += (((regs.eax >> 16) & 0x0Fu) << 4);
    }

    if (family_id == 0x0Fu) {
      family_id += ((regs.eax >> 20) & 0xFFu);
    }

    cpu._model_id = model_id;
    cpu._family_id = family_id;
    cpu._brand_id = (regs.ebx) & 0xFF;
    cpu._processor_type = (regs.eax >> 12) & 0x03;
    cpu._max_logical_processors = (regs.ebx >> 16) & 0xFF;
    cpu._stepping = (regs.eax) & 0x0F;
    cpu._cache_line_size  = ((regs.ebx >> 8) & 0xFF) * 8;

    features.add_if(bit_test(regs.ecx,  0), Ext::kSSE3);
    features.add_if(bit_test(regs.ecx,  1), Ext::kPCLMULQDQ);
    features.add_if(bit_test(regs.ecx,  3), Ext::kMONITOR);
    features.add_if(bit_test(regs.ecx,  5), Ext::kVMX);
    features.add_if(bit_test(regs.ecx,  6), Ext::kSMX);
    features.add_if(bit_test(regs.ecx,  9), Ext::kSSSE3);
    features.add_if(bit_test(regs.ecx, 13), Ext::kCMPXCHG16B);
    features.add_if(bit_test(regs.ecx, 19), Ext::kSSE4_1);
    features.add_if(bit_test(regs.ecx, 20), Ext::kSSE4_2);
    features.add_if(bit_test(regs.ecx, 22), Ext::kMOVBE);
    features.add_if(bit_test(regs.ecx, 23), Ext::kPOPCNT);
    features.add_if(bit_test(regs.ecx, 25), Ext::kAESNI);
    features.add_if(bit_test(regs.ecx, 26), Ext::kXSAVE);
    features.add_if(bit_test(regs.ecx, 27), Ext::kOSXSAVE);
    features.add_if(bit_test(regs.ecx, 30), Ext::kRDRAND);
    features.add_if(bit_test(regs.edx,  0), Ext::kFPU);
    features.add_if(bit_test(regs.edx,  4), Ext::kRDTSC);
    features.add_if(bit_test(regs.edx,  5), Ext::kMSR);
    features.add_if(bit_test(regs.edx,  8), Ext::kCMPXCHG8B);
    features.add_if(bit_test(regs.edx, 15), Ext::kCMOV);
    features.add_if(bit_test(regs.edx, 19), Ext::kCLFLUSH);
    features.add_if(bit_test(regs.edx, 23), Ext::kMMX);
    features.add_if(bit_test(regs.edx, 24), Ext::kFXSR);
    features.add_if(bit_test(regs.edx, 25), Ext::kSSE, Ext::kMMX2);
    features.add_if(bit_test(regs.edx, 26), Ext::kSSE2, Ext::kSSE);
    features.add_if(bit_test(regs.edx, 28), Ext::kMT);

    // Get the content of XCR0 if supported by the CPU and enabled by the OS.
    if (features.has_xsave() && features.has_osxsave()) {
      xgetbv_query(&xcr0, 0);
    }

    // Detect AVX+.
    if (bit_test(regs.ecx, 28)) {
      // - XCR0[2:1] == 11b
      //   XMM & YMM states need to be enabled by OS.
      if ((xcr0.eax & 0x00000006u) == 0x00000006u) {
        features.add(Ext::kAVX);
        features.add_if(bit_test(regs.ecx, 12), Ext::kFMA);
        features.add_if(bit_test(regs.ecx, 29), Ext::kF16C);
      }
    }
  }

  constexpr uint32_t kXCR0_AMX_Bits = 0x3u << 17;
  bool amx_enabled = (xcr0.eax & kXCR0_AMX_Bits) == kXCR0_AMX_Bits;

#if defined(__APPLE__)
  // Apple platform provides on-demand AVX512 support. When an AVX512 instruction is used the first time it results
  // in #UD, which would cause the thread being promoted to use AVX512 support by the OS in addition to enabling the
  // necessary bits in XCR0 register.
  bool avx512_enabled = true;
#else
  // - XCR0[2:1] ==  11b - XMM/YMM states need to be enabled by OS.
  // - XCR0[7:5] == 111b - Upper 256-bit of ZMM0-XMM15 and ZMM16-ZMM31 need to be enabled by OS.
  constexpr uint32_t kXCR0_AVX512_Bits = (0x3u << 1) | (0x7u << 5);
  bool avx512_enabled = (xcr0.eax & kXCR0_AVX512_Bits) == kXCR0_AVX512_Bits;
#endif

  bool avx10_enabled = false;

  // CPUID EAX=0x07 ECX=0 (Structured Extended Feature Flags Enumeration Leaf)
  // -------------------------------------------------------------------------

  if (max_id >= 0x07u) {
    cpuid_query(&regs, 0x7);

    max_sub_leaf_id_0x7 = regs.eax;

    features.add_if(bit_test(regs.ebx,  0), Ext::kFSGSBASE);
    features.add_if(bit_test(regs.ebx,  3), Ext::kBMI);
    features.add_if(bit_test(regs.ebx,  7), Ext::kSMEP);
    features.add_if(bit_test(regs.ebx,  8), Ext::kBMI2);
    features.add_if(bit_test(regs.ebx,  9), Ext::kERMS);
    features.add_if(bit_test(regs.ebx, 18), Ext::kRDSEED);
    features.add_if(bit_test(regs.ebx, 19), Ext::kADX);
    features.add_if(bit_test(regs.ebx, 20), Ext::kSMAP);
    features.add_if(bit_test(regs.ebx, 23), Ext::kCLFLUSHOPT);
    features.add_if(bit_test(regs.ebx, 24), Ext::kCLWB);
    features.add_if(bit_test(regs.ebx, 29), Ext::kSHA);
    features.add_if(bit_test(regs.ecx,  0), Ext::kPREFETCHWT1);
    features.add_if(bit_test(regs.ecx,  4), Ext::kOSPKE);
    features.add_if(bit_test(regs.ecx,  5), Ext::kWAITPKG);
    features.add_if(bit_test(regs.ecx,  7), Ext::kCET_SS);
    features.add_if(bit_test(regs.ecx,  8), Ext::kGFNI);
    features.add_if(bit_test(regs.ecx,  9), Ext::kVAES);
    features.add_if(bit_test(regs.ecx, 10), Ext::kVPCLMULQDQ);
    features.add_if(bit_test(regs.ecx, 22), Ext::kRDPID);
    features.add_if(bit_test(regs.ecx, 23), Ext::kKL);
    features.add_if(bit_test(regs.ecx, 25), Ext::kCLDEMOTE);
    features.add_if(bit_test(regs.ecx, 27), Ext::kMOVDIRI);
    features.add_if(bit_test(regs.ecx, 28), Ext::kMOVDIR64B);
    features.add_if(bit_test(regs.ecx, 29), Ext::kENQCMD);
    features.add_if(bit_test(regs.edx,  4), Ext::kFSRM);
    features.add_if(bit_test(regs.edx,  5), Ext::kUINTR);
    features.add_if(bit_test(regs.edx, 14), Ext::kSERIALIZE);
    features.add_if(bit_test(regs.edx, 16), Ext::kTSXLDTRK);
    features.add_if(bit_test(regs.edx, 18), Ext::kPCONFIG);
    features.add_if(bit_test(regs.edx, 20), Ext::kCET_IBT);

    if (bit_test(regs.ebx, 5) && features.has_avx()) {
      features.add(Ext::kAVX2);
    }

    if (avx512_enabled && bit_test(regs.ebx, 16)) {
      features.add(Ext::kAVX512_F);

      features.add_if(bit_test(regs.ebx, 17), Ext::kAVX512_DQ);
      features.add_if(bit_test(regs.ebx, 21), Ext::kAVX512_IFMA);
      features.add_if(bit_test(regs.ebx, 28), Ext::kAVX512_CD);
      features.add_if(bit_test(regs.ebx, 30), Ext::kAVX512_BW);
      features.add_if(bit_test(regs.ebx, 31), Ext::kAVX512_VL);
      features.add_if(bit_test(regs.ecx,  1), Ext::kAVX512_VBMI);
      features.add_if(bit_test(regs.ecx,  6), Ext::kAVX512_VBMI2);
      features.add_if(bit_test(regs.ecx, 11), Ext::kAVX512_VNNI);
      features.add_if(bit_test(regs.ecx, 12), Ext::kAVX512_BITALG);
      features.add_if(bit_test(regs.ecx, 14), Ext::kAVX512_VPOPCNTDQ);
      features.add_if(bit_test(regs.edx,  8), Ext::kAVX512_VP2INTERSECT);
      features.add_if(bit_test(regs.edx, 23), Ext::kAVX512_FP16);
    }

    if (amx_enabled) {
      features.add_if(bit_test(regs.edx, 22), Ext::kAMX_BF16);
      features.add_if(bit_test(regs.edx, 24), Ext::kAMX_TILE);
      features.add_if(bit_test(regs.edx, 25), Ext::kAMX_INT8);
    }
  }

  // CPUID EAX=0x07 ECX=1 (Structured Extended Feature Enumeration Sub-leaf)
  // -----------------------------------------------------------------------

  if (max_sub_leaf_id_0x7 >= 1) {
    cpuid_query(&regs, 0x7, 1);

    features.add_if(bit_test(regs.eax,  0), Ext::kSHA512);
    features.add_if(bit_test(regs.eax,  1), Ext::kSM3);
    features.add_if(bit_test(regs.eax,  2), Ext::kSM4);
    features.add_if(bit_test(regs.eax,  3), Ext::kRAO_INT);
    features.add_if(bit_test(regs.eax,  7), Ext::kCMPCCXADD);
    features.add_if(bit_test(regs.eax, 10), Ext::kFZRM);
    features.add_if(bit_test(regs.eax, 11), Ext::kFSRS);
    features.add_if(bit_test(regs.eax, 12), Ext::kFSRC);
    features.add_if(bit_test(regs.eax, 19), Ext::kWRMSRNS);
    features.add_if(bit_test(regs.eax, 22), Ext::kHRESET);
    features.add_if(bit_test(regs.eax, 26), Ext::kLAM);
    features.add_if(bit_test(regs.eax, 27), Ext::kMSRLIST);
    features.add_if(bit_test(regs.eax, 31), Ext::kMOVRS);
    features.add_if(bit_test(regs.ecx,  5), Ext::kMSR_IMM);
    features.add_if(bit_test(regs.ebx,  1), Ext::kTSE);
    features.add_if(bit_test(regs.edx, 14), Ext::kPREFETCHI);
    features.add_if(bit_test(regs.edx, 18), Ext::kCET_SSS);
    features.add_if(bit_test(regs.edx, 21), Ext::kAPX_F);

    if (features.has_avx2()) {
      features.add_if(bit_test(regs.eax,  4), Ext::kAVX_VNNI);
      features.add_if(bit_test(regs.eax, 23), Ext::kAVX_IFMA);
      features.add_if(bit_test(regs.edx,  4), Ext::kAVX_VNNI_INT8);
      features.add_if(bit_test(regs.edx,  5), Ext::kAVX_NE_CONVERT);
      features.add_if(bit_test(regs.edx, 10), Ext::kAVX_VNNI_INT16);
    }

    if (features.has_avx512_f()) {
      features.add_if(bit_test(regs.eax,  5), Ext::kAVX512_BF16);
    }

    if (features.has_avx512_f()) {
      avx10_enabled = Support::bit_test(regs.edx, 19);
    }

    if (amx_enabled) {
      features.add_if(bit_test(regs.eax, 21), Ext::kAMX_FP16);
      features.add_if(bit_test(regs.edx,  8), Ext::kAMX_COMPLEX);
    }
  }

  // CPUID EAX=0x0D ECX=1 (Processor Extended State Enumeration Sub-leaf)
  // --------------------------------------------------------------------

  if (max_id >= 0x0Du) {
    cpuid_query(&regs, 0xD, 1);

    features.add_if(bit_test(regs.eax, 0), Ext::kXSAVEOPT);
    features.add_if(bit_test(regs.eax, 1), Ext::kXSAVEC);
    features.add_if(bit_test(regs.eax, 3), Ext::kXSAVES);
  }

  // CPUID EAX=0x0E ECX=0 (Processor Trace Enumeration Main Leaf)
  // ------------------------------------------------------------

  if (max_id >= 0x0Eu) {
    cpuid_query(&regs, 0x0E, 0);

    features.add_if(bit_test(regs.ebx, 4), Ext::kPTWRITE);
  }

  // CPUID EAX=0x19 ECX=0 (Key Locker Leaf)
  // --------------------------------------

  if (max_id >= 0x19u && features.has_kl()) {
    cpuid_query(&regs, 0x19, 0);

    features.add_if(bit_test(regs.ebx, 0), Ext::kAESKLE);
    features.add_if(bit_test(regs.ebx, 0) && bit_test(regs.ebx, 2), Ext::kAESKLEWIDE_KL);
  }

  // CPUID EAX=0x1E ECX=1 (TMUL Information Sub-leaf)
  // ------------------------------------------------

  if (max_id >= 0x1Eu && features.has_amx_tile()) {
    cpuid_query(&regs, 0x1E, 1);

    // NOTE: Some AMX flags are mirrored here from CPUID[0x07, 0x00].
    features.add_if(bit_test(regs.eax, 0), Ext::kAMX_INT8);
    features.add_if(bit_test(regs.eax, 1), Ext::kAMX_BF16);
    features.add_if(bit_test(regs.eax, 2), Ext::kAMX_COMPLEX);
    features.add_if(bit_test(regs.eax, 3), Ext::kAMX_FP16);
    features.add_if(bit_test(regs.eax, 4), Ext::kAMX_FP8);
    features.add_if(bit_test(regs.eax, 5), Ext::kAMX_TRANSPOSE);
    features.add_if(bit_test(regs.eax, 6), Ext::kAMX_TF32);
    features.add_if(bit_test(regs.eax, 7), Ext::kAMX_AVX512);
    features.add_if(bit_test(regs.eax, 8), Ext::kAMX_MOVRS);
  }

  // CPUID EAX=0x24 ECX=0 (AVX10 Information)
  // ----------------------------------------

  if (max_id >= 0x24u && avx10_enabled) {
    // EAX output is the maximum supported sub-leaf.
    cpuid_query(&regs, 0x24, 0);

    // AVX10 Converged Vector ISA version.
    uint32_t ver = regs.ebx & 0xFFu;

    features.add_if(ver >= 1u, Ext::kAVX10_1);
    features.add_if(ver >= 2u, Ext::kAVX10_2);
  }

  // CPUID EAX=0x80000000...max_id
  // ----------------------------

  max_id = 0x80000000u;
  uint32_t i = max_id;

  // The highest EAX that we understand.
  constexpr uint32_t kHighestProcessedEAX = 0x8000001Fu;

  // Several CPUID calls are required to get the whole brand string. It's easier
  // to copy one DWORD at a time instead of copying the string a byte by byte.
  uint32_t* brand = cpu._brand.u32;
  do {
    cpuid_query(&regs, i);
    switch (i) {
      case 0x80000000u:
        max_id = Support::min<uint32_t>(regs.eax, kHighestProcessedEAX);
        break;

      case 0x80000001u:
        features.add_if(bit_test(regs.ecx,  0), Ext::kLAHFSAHF);
        features.add_if(bit_test(regs.ecx,  2), Ext::kSVM);
        features.add_if(bit_test(regs.ecx,  5), Ext::kLZCNT);
        features.add_if(bit_test(regs.ecx,  6), Ext::kSSE4A);
        features.add_if(bit_test(regs.ecx,  7), Ext::kMSSE);
        features.add_if(bit_test(regs.ecx,  8), Ext::kPREFETCHW);
        features.add_if(bit_test(regs.ecx, 12), Ext::kSKINIT);
        features.add_if(bit_test(regs.ecx, 15), Ext::kLWP);
        features.add_if(bit_test(regs.ecx, 21), Ext::kTBM);
        features.add_if(bit_test(regs.ecx, 29), Ext::kMONITORX);
        features.add_if(bit_test(regs.edx, 20), Ext::kNX);
        features.add_if(bit_test(regs.edx, 21), Ext::kFXSROPT);
        features.add_if(bit_test(regs.edx, 22), Ext::kMMX2);
        features.add_if(bit_test(regs.edx, 27), Ext::kRDTSCP);
        features.add_if(bit_test(regs.edx, 29), Ext::kPREFETCHW);
        features.add_if(bit_test(regs.edx, 30), Ext::k3DNOW2, Ext::kMMX2);
        features.add_if(bit_test(regs.edx, 31), Ext::kPREFETCHW);

        if (features.has_avx()) {
          features.add_if(bit_test(regs.ecx, 11), Ext::kXOP);
          features.add_if(bit_test(regs.ecx, 16), Ext::kFMA4);
        }

        // This feature seems to be only supported by AMD.
        if (cpu.is_vendor("AMD")) {
          features.add_if(bit_test(regs.ecx,  4), Ext::kALTMOVCR8);
        }
        break;

      case 0x80000002u:
      case 0x80000003u:
      case 0x80000004u:
        *brand++ = regs.eax;
        *brand++ = regs.ebx;
        *brand++ = regs.ecx;
        *brand++ = regs.edx;

        // Go directly to the next one we are interested in.
        if (i == 0x80000004u)
          i = 0x80000008u - 1;
        break;

      case 0x80000008u:
        features.add_if(bit_test(regs.ebx,  0), Ext::kCLZERO);
        features.add_if(bit_test(regs.ebx,  0), Ext::kRDPRU);
        features.add_if(bit_test(regs.ebx,  8), Ext::kMCOMMIT);
        features.add_if(bit_test(regs.ebx,  9), Ext::kWBNOINVD);

        // Go directly to the next one we are interested in.
        i = 0x8000001Fu - 1;
        break;

      case 0x8000001Fu:
        features.add_if(bit_test(regs.eax,  0), Ext::kSME);
        features.add_if(bit_test(regs.eax,  1), Ext::kSEV);
        features.add_if(bit_test(regs.eax,  3), Ext::kSEV_ES);
        features.add_if(bit_test(regs.eax,  4), Ext::kSEV_SNP);
        features.add_if(bit_test(regs.eax,  6), Ext::kRMPQUERY);
        break;
    }
  } while (++i <= max_id);

  // Simplify CPU brand string a bit by removing some unnecessary spaces.
  simplify_cpu_brand(cpu._brand.str);
}

static ASMJIT_FAVOR_SIZE CpuHints recalculate_hints(const CpuInfo& cpu_info, const CpuFeatures::X86& features) noexcept {
  CpuHints hints {};

  // Vendor Independent CPU Hints
  // ----------------------------

  if (features.has_avx2()) {
    hints |= CpuHints::kVecMaskedOps32 | CpuHints::kVecMaskedOps64;
  }

  if (features.has_avx512_bw()) {
    hints |= CpuHints::kVecMaskedOps8 | CpuHints::kVecMaskedOps16 | CpuHints::kVecMaskedOps32 | CpuHints::kVecMaskedOps64;
  }

  // Select optimization flags based on CPU vendor and micro-architecture.

  // AMD Specific CPU Hints
  // ----------------------

  if (cpu_info.is_vendor("AMD")) {
    // Zen 3+ has fast gathers, scalar loads and shuffles are faster on Zen 2 and older CPUs.
    if (cpu_info.family_id() >= 0x19u) {
      hints |= CpuHints::kVecFastGather;
    }

    // Zen 1+ provides low-latency VPMULLD instruction.
    if (features.has_avx2()) {
      hints |= CpuHints::kVecFastIntMul32;
    }

    // Zen 4+ provides low-latency VPMULLQ instruction.
    if (features.has_avx512_dq()) {
      hints |= CpuHints::kVecFastIntMul64;
    }

    // Zen 4+ has fast mask operations (starts with AVX-512).
    if (features.has_avx512_f()) {
      hints |= CpuHints::kVecMaskedStore;
    }
  }

  // Intel Specific CPU Hints
  // ------------------------

  if (cpu_info.is_vendor("INTEL")) {
    if (features.has_avx2()) {
      uint32_t family_id = cpu_info.family_id();
      uint32_t model_id = cpu_info.model_id();

      // NOTE: We only want to hint fast gathers in cases the CPU is immune to DOWNFALL. The reason is that the
      // DOWNFALL mitigation delivered via a micro-code update makes gathers almost useless in a way that scalar
      // loads can beat it significantly (in Blend2D case scalar loads can offer up to 50% more performance).
      // This table basically picks CPUs that are known to not be affected by DOWNFALL.
      if (family_id == 0x06u) {
        switch (model_id) {
          case 0x8Fu: // Sapphire Rapids.
          case 0x96u: // Elkhart Lake.
          case 0x97u: // Alder Lake / Catlow.
          case 0x9Au: // Alder Lake / Arizona Beach.
          case 0x9Cu: // Jasper Lake.
          case 0xAAu: // Meteor Lake.
          case 0xACu: // Meteor Lake.
          case 0xADu: // Granite Rapids.
          case 0xAEu: // Granite Rapids.
          case 0xAFu: // Sierra Forest.
          case 0xBAu: // Raptor Lake.
          case 0xB5u: // Arrow Lake.
          case 0xB6u: // Grand Ridge.
          case 0xB7u: // Raptor Lake / Catlow.
          case 0xBDu: // Lunar Lake.
          case 0xBEu: // Alder Lake (N).
          case 0xBFu: // Raptor Lake.
          case 0xC5u: // Arrow Lake.
          case 0xC6u: // Arrow Lake.
          case 0xCFu: // Emerald Rapids.
          case 0xDDu: // Clearwater Forest.
            hints |= CpuHints::kVecFastGather;
            break;

          default:
            break;
        }
      }
    }

    // TODO: It seems masked stores are very expensive on consumer INTEL CPUs.
    // hints |= CpuHints::kVecMaskedStore;
  }

  return hints;
}

} // {x86}

#endif // ASMJIT_ARCH_X86

// CpuInfo - Detect - ARM
// ======================

// Implement the most code outside the platform specific #ifdefs to minimize breaking the detection on
// platforms that don't run on our CI infrastructure. The problem with the detection is that every OS
// requires a specific implementation as ARM features cannot be detected in user-mode without OS enablement.

// The most relevant and accurate information can be found here:
//   https://github.com/llvm-project/llvm/blob/master/lib/Target/AArch64/AArch64.td
//   https://github.com/apple/llvm-project/blob/apple/main/llvm/lib/Target/AArch64/AArch64.td (Apple fork)
//
// Other resources:
//   https://en.wikipedia.org/wiki/AArch64
//   https://en.wikipedia.org/wiki/Apple_silicon#List_of_Apple_processors
//   https://developer.arm.com/downloads/-/exploration-tools/feature-names-for-a-profile
//   https://developer.arm.com/architectures/learn-the-architecture/understanding-the-armv8-x-extensions/single-page

#if ASMJIT_ARCH_ARM

namespace arm {

// ARM commonly refers to CPU features using FEAT_ prefix, we use Ext:: to make it compatible with other parts.
using Ext = CpuFeatures::ARM;

// CpuInfo - Detect - ARM - OS Kernel Version
// ==========================================

#if defined(__linux__)
struct UNameKernelVersion {
  int parts[3];

  inline bool at_least(int major, int minor, int patch = 0) const noexcept {
    if (parts[0] >= major) {
      if (parts[0] > major) {
        return true;
      }

      if (parts[1] >= minor) {
        return parts[1] > minor ? true : parts[2] >= patch;
      }
    }

    return false;
  }
};

[[maybe_unused]]
static UNameKernelVersion get_kernel_version_via_uname() noexcept {
  UNameKernelVersion ver{};
  ver.parts[0] = -1;

  utsname buffer;
  if (uname(&buffer) != 0) {
    return ver;
  }

  size_t count = 0;
  char* p = buffer.release;
  while (*p) {
    uint32_t c = uint8_t(*p);
    if (c >= uint32_t('0') && c <= uint32_t('9')) {
      ver.parts[count] = int(strtol(p, &p, 10));
      if (++count == 3) {
        break;
      }
    }
    else if (c == '.' || c == '-') {
      p++;
    }
    else {
      break;
    }
  }

  return ver;
}
#endif // __linux__

// CpuInfo - Detect - ARM - Baseline Features of ARM Architectures
// ===============================================================

[[maybe_unused]]
static inline void populate_base_aarch32_features(CpuFeatures::ARM& features) noexcept {
  // No baseline flags at the moment.
  Support::maybe_unused(features);
}

[[maybe_unused]]
static inline void populate_base_aarch64_features(CpuFeatures::ARM& features) noexcept {
  // AArch64 is based on ARMv8.0 and later.
  features.add(Ext::kARMv6);
  features.add(Ext::kARMv7);
  features.add(Ext::kARMv8a);

  // AArch64 comes with these features by default.
  features.add(Ext::kASIMD);
  features.add(Ext::kFP);
  features.add(Ext::kIDIVA);
}

static inline void populate_base_arm_features(CpuInfo& cpu) noexcept {
#if ASMJIT_ARCH_ARM == 32
  populate_base_aarch32_features(cpu.features().arm());
#else
  populate_base_aarch64_features(cpu.features().arm());
#endif
}

// CpuInfo - Detect - ARM - Mandatory Features of ARM Architectures
// ================================================================

// Populates mandatory ARMv8.[v]A features.
[[maybe_unused]]
static ASMJIT_NOINLINE void populate_armv8a_features(CpuFeatures::ARM& features, uint32_t v) noexcept {
  switch (v) {
    default:
      [[fallthrough]];
    case 9: // ARMv8.9
      features.add(Ext::kCLRBHB, Ext::kCSSC, Ext::kPRFMSLC, Ext::kSPECRES2, Ext::kRAS2);
      [[fallthrough]];
    case 8: // ARMv8.8
      features.add(Ext::kHBC, Ext::kMOPS, Ext::kNMI);
      [[fallthrough]];
    case 7: // ARMv8.7
      features.add(Ext::kHCX, Ext::kPAN3, Ext::kWFXT, Ext::kXS);
      [[fallthrough]];
    case 6: // ARMv8.6
      features.add(Ext::kAMU1_1, Ext::kBF16, Ext::kECV, Ext::kFGT, Ext::kI8MM);
      [[fallthrough]];
    case 5: // ARMv8.5
      features.add(Ext::kBTI, Ext::kCSV2, Ext::kDPB2, Ext::kFLAGM2, Ext::kFRINTTS, Ext::kSB, Ext::kSPECRES, Ext::kSSBS);
      [[fallthrough]];
    case 4: // ARMv8.4
      features.add(Ext::kAMU1, Ext::kDIT, Ext::kDOTPROD, Ext::kFLAGM,
                   Ext::kLRCPC2, Ext::kLSE2, Ext::kMPAM, Ext::kNV,
                   Ext::kSEL2, Ext::kTLBIOS, Ext::kTLBIRANGE, Ext::kTRF);
      [[fallthrough]];
    case 3: // ARMv8.3
      features.add(Ext::kCCIDX, Ext::kFCMA, Ext::kJSCVT, Ext::kLRCPC, Ext::kPAUTH);
      [[fallthrough]];
    case 2: // ARMv8.2
      features.add(Ext::kDPB, Ext::kPAN2, Ext::kRAS, Ext::kUAO);
      [[fallthrough]];
    case 1: // ARMv8.1
      features.add(Ext::kCRC32, Ext::kLOR, Ext::kLSE, Ext::kPAN, Ext::kRDM, Ext::kVHE);
      [[fallthrough]];
    case 0: // ARMv8.0
      features.add(Ext::kASIMD, Ext::kFP, Ext::kIDIVA, Ext::kVFP_D32);
      break;
  }
}

// Populates mandatory ARMv9.[v] features.
[[maybe_unused]]
static ASMJIT_FAVOR_SIZE void populate_armv9a_features(CpuFeatures::ARM& features, uint32_t v) noexcept {
  populate_armv8a_features(features, v <= 4u ? 5u + v : 9u);

  switch (v) {
    default:
      [[fallthrough]];
    case 4: // ARMv9.4 - based on ARMv8.9.
      [[fallthrough]];
    case 3: // ARMv9.3 - based on ARMv8.8.
      [[fallthrough]];
    case 2: // ARMv9.2 - based on ARMv8.7.
      [[fallthrough]];
    case 1: // ARMv9.1 - based on ARMv8.6.
      [[fallthrough]];
    case 0: // ARMv9.0 - based on ARMv8.5.
      features.add(Ext::kRME, Ext::kSVE, Ext::kSVE2);
      break;
  }
}

// CpuInfo - Detect - ARM - CPUID Based Features
// =============================================

// This implements detection based on the content of CPUID registers. The following code doesn't actually read any
// of the registers so it's an implementation that can theoretically be tested / used in mocks.

// Merges a feature that contains 0b1111 when it doesn't exist and starts at 0b0000 when it does.
[[maybe_unused]]
static ASMJIT_INLINE void merge_aarch64_cpuid_feature_na(
  CpuFeatures::ARM& features, uint64_t reg_bits, uint32_t offset,
  Ext::Id f0,
  Ext::Id f1 = Ext::kNone,
  Ext::Id f2 = Ext::kNone,
  Ext::Id f3 = Ext::kNone) noexcept {

  uint32_t val = uint32_t((reg_bits >> offset) & 0xFu);
  if (val == 0xFu) {
    // If val == 0b1111 then the feature is not implemented in this case (some early extensions).
    return;
  }

  features.add_if(f0 != Ext::kNone, f0);
  features.add_if(f1 != Ext::kNone && val >= 1, f1);
  features.add_if(f2 != Ext::kNone && val >= 2, f2);
  features.add_if(f3 != Ext::kNone && val >= 3, f3);
}

// Merges a feature identified by a single bit at `offset`.
[[maybe_unused]]
static ASMJIT_INLINE void merge_aarch64_cpuid_feature_1b(CpuFeatures::ARM& features, uint64_t reg_bits, uint32_t offset, Ext::Id f1) noexcept {
  features.add_if((reg_bits & (uint64_t(1) << offset)) != 0, f1);
}

// Merges a feature-list starting from 0b01 when it does (0b00 means feature not supported).
[[maybe_unused]]
static ASMJIT_INLINE void merge_aarch64_cpuid_feature_2b(CpuFeatures::ARM& features, uint64_t reg_bits, uint32_t offset, Ext::Id f1, Ext::Id f2, Ext::Id f3) noexcept {
  uint32_t val = uint32_t((reg_bits >> offset) & 0x3u);

  features.add_if(f1 != Ext::kNone && val >= 1, f1);
  features.add_if(f2 != Ext::kNone && val >= 2, f2);
  features.add_if(f3 != Ext::kNone && val == 3, f3);
}

// Merges a feature-list starting from 0b0001 when it does (0b0000 means feature not supported).
[[maybe_unused]]
static ASMJIT_INLINE void merge_aarch64_cpuid_feature_4b(CpuFeatures::ARM& features, uint64_t reg_bits, uint32_t offset,
  Ext::Id f1,
  Ext::Id f2 = Ext::kNone,
  Ext::Id f3 = Ext::kNone,
  Ext::Id f4 = Ext::kNone) noexcept {

  uint32_t val = uint32_t((reg_bits >> offset) & 0xFu);

  // if val == 0 it means that this feature is not supported.
  features.add_if(f1 != Ext::kNone && val >= 1, f1);
  features.add_if(f2 != Ext::kNone && val >= 2, f2);
  features.add_if(f3 != Ext::kNone && val >= 3, f3);
  features.add_if(f4 != Ext::kNone && val >= 4, f4);
}

// Merges a feature that is identified by an exact bit-combination of 4 bits.
[[maybe_unused]]
static ASMJIT_INLINE void merge_aarch64_cpuid_feature_4s(CpuFeatures::ARM& features, uint64_t reg_bits, uint32_t offset, uint32_t value, Ext::Id f1) noexcept {
  features.add_if(uint32_t((reg_bits >> offset) & 0xFu) == value, f1);
}

#define MERGE_FEATURE_NA(identifier, reg, offset, ...) merge_aarch64_cpuid_feature_na(cpu.features().arm(), reg, offset, __VA_ARGS__)
#define MERGE_FEATURE_1B(identifier, reg, offset, ...) merge_aarch64_cpuid_feature_1b(cpu.features().arm(), reg, offset, __VA_ARGS__)
#define MERGE_FEATURE_2B(identifier, reg, offset, ...) merge_aarch64_cpuid_feature_2b(cpu.features().arm(), reg, offset, __VA_ARGS__)
#define MERGE_FEATURE_4B(identifier, reg, offset, ...) merge_aarch64_cpuid_feature_4b(cpu.features().arm(), reg, offset, __VA_ARGS__)
#define MERGE_FEATURE_4S(identifier, reg, offset, ...) merge_aarch64_cpuid_feature_4s(cpu.features().arm(), reg, offset, __VA_ARGS__)

// Detects features based on the content of ID_AA64PFR0_EL1 and ID_AA64PFR1_EL1 registers.
[[maybe_unused]]
static inline void detect_aarch64_features_via_cpuid_aa64pfr0_aa64pfr1(CpuInfo& cpu, uint64_t fpr0, uint64_t fpr1) noexcept {
  // ID_AA64PFR0_EL1
  // ===============

  // FP and AdvSIMD bits should match (i.e. if FP features FP16, ASIMD must feature it too).
  MERGE_FEATURE_NA("FP bits [19:16]"          , fpr0, 16, Ext::kFP, Ext::kFP16);
  MERGE_FEATURE_NA("AdvSIMD bits [23:20]"     , fpr0, 20, Ext::kASIMD, Ext::kFP16);
  /*
  MERGE_FEATURE_4B("GIC bits [27:24]"         , fpr0, 24, ...);
  */
  MERGE_FEATURE_4B("RAS bits [31:28]"         , fpr0, 28, Ext::kRAS, Ext::kRAS1_1, Ext::kRAS2);
  MERGE_FEATURE_4B("SVE bits [35:32]"         , fpr0, 32, Ext::kSVE);
  MERGE_FEATURE_4B("SEL2 bits [39:36]"        , fpr0, 36, Ext::kSEL2);
  MERGE_FEATURE_4B("MPAM bits [43:40]"        , fpr0, 40, Ext::kMPAM);
  MERGE_FEATURE_4B("AMU bits [47:44]"         , fpr0, 44, Ext::kAMU1, Ext::kAMU1_1);
  MERGE_FEATURE_4B("DIT bits [51:48]"         , fpr0, 48, Ext::kDIT);
  MERGE_FEATURE_4B("RME bits [55:52]"         , fpr0, 52, Ext::kRME);
  MERGE_FEATURE_4B("CSV2 bits [59:56]"        , fpr0, 56, Ext::kCSV2, Ext::kCSV2, Ext::kCSV2, Ext::kCSV2_3);
  MERGE_FEATURE_4B("CSV3 bits [63:60]"        , fpr0, 60, Ext::kCSV3);

  // ID_AA64PFR1_EL1
  // ===============

  MERGE_FEATURE_4B("BT bits [3:0]"            , fpr1,  0, Ext::kBTI);
  MERGE_FEATURE_4B("SSBS bits [7:4]"          , fpr1,  4, Ext::kSSBS, Ext::kSSBS2);
  MERGE_FEATURE_4B("MTE bits [11:8]"          , fpr1,  8, Ext::kMTE, Ext::kMTE2, Ext::kMTE3);
  /*
  MERGE_FEATURE_4B("RAS_frac bits [15:12]"    , fpr1, 12, ...);
  MERGE_FEATURE_4B("MPAM_frac bits [19:16]"   , fpr1, 16, ...);
  */
  MERGE_FEATURE_4B("SME bits [27:24]"         , fpr1, 24, Ext::kSME, Ext::kSME2);
  MERGE_FEATURE_4B("RNDR_trap bits [31:28]"   , fpr1, 28, Ext::kRNG_TRAP);
  /*
  MERGE_FEATURE_4B("CSV2_frac bits [35:32]"   , fpr1, 32, ...);
  */
  MERGE_FEATURE_4B("NMI bits [39:36]"         , fpr1, 36, Ext::kNMI);
  /*
  MERGE_FEATURE_4B("MTE_frac bits [43:40]"    , fpr1, 40, ...);
  */
  MERGE_FEATURE_4B("GCS bits [47:44]"         , fpr1, 44, Ext::kGCS);
  MERGE_FEATURE_4B("THE bits [51:48]"         , fpr1, 48, Ext::kTHE);

  // MTEX extensions are only available when MTE3 is available.
  if (cpu.features().arm().has_mte3())
    MERGE_FEATURE_4B("MTEX bits [55:52]"      , fpr1, 52, Ext::kMTE4);

  /*
  MERGE_FEATURE_4B("DF2 bits [59:56]"         , fpr1, 56, ...);
  */
  MERGE_FEATURE_4B("PFAR bits [63:60]"        , fpr1, 60, Ext::kPFAR);

  // ID_AA64PFR0_EL1 + ID_AA64PFR1_EL1
  // =================================

  uint32_t ras_main = uint32_t((fpr0 >> 28) & 0xFu);
  uint32_t ras_frac = uint32_t((fpr1 >> 12) & 0xFu);

  if (ras_main == 1 && ras_frac == 1) {
    cpu.features().arm().add(Ext::kRAS1_1);
  }

  uint32_t mpam_main = uint32_t((fpr0 >> 40) & 0xFu);
  uint32_t mpam_frac = uint32_t((fpr1 >> 16) & 0xFu);

  if (mpam_main || mpam_frac) {
    cpu.features().arm().add(Ext::kMPAM);
  }
}

// Detects features based on the content of ID_AA64ISAR0_EL1 and ID_AA64ISAR1_EL1 registers.
[[maybe_unused]]
static inline void detect_aarch64_features_via_cpuid_aa64isar0_aa64isar1(CpuInfo& cpu, uint64_t isar0, uint64_t isar1) noexcept {
  // ID_AA64ISAR0_EL1
  // ================

  MERGE_FEATURE_4B("AES bits [7:4]"           , isar0,  4, Ext::kAES, Ext::kPMULL);
  MERGE_FEATURE_4B("SHA1 bits [11:8]"         , isar0,  8, Ext::kSHA1);
  MERGE_FEATURE_4B("SHA2 bits [15:12]"        , isar0, 12, Ext::kSHA256, Ext::kSHA512);
  MERGE_FEATURE_4B("CRC32 bits [19:16]"       , isar0, 16, Ext::kCRC32);
  MERGE_FEATURE_4B("Atomic bits [23:20]"      , isar0, 20, Ext::kNone, Ext::kLSE, Ext::kLSE128);
  MERGE_FEATURE_4B("TME bits [27:24]"         , isar0, 24, Ext::kTME);
  MERGE_FEATURE_4B("RDM bits [31:28]"         , isar0, 28, Ext::kRDM);
  MERGE_FEATURE_4B("SHA3 bits [35:32]"        , isar0, 32, Ext::kSHA3);
  MERGE_FEATURE_4B("SM3 bits [39:36]"         , isar0, 36, Ext::kSM3);
  MERGE_FEATURE_4B("SM4 bits [43:40]"         , isar0, 40, Ext::kSM4);
  MERGE_FEATURE_4B("DP bits [47:44]"          , isar0, 44, Ext::kDOTPROD);
  MERGE_FEATURE_4B("FHM bits [51:48]"         , isar0, 48, Ext::kFHM);
  MERGE_FEATURE_4B("TS bits [55:52]"          , isar0, 52, Ext::kFLAGM, Ext::kFLAGM2);
  /*
  MERGE_FEATURE_4B("TLB bits [59:56]"         , isar0, 56, ...);
  */
  MERGE_FEATURE_4B("RNDR bits [63:60]"        , isar0, 60, Ext::kFLAGM, Ext::kRNG);

  // ID_AA64ISAR1_EL1
  // ================

  MERGE_FEATURE_4B("DPB bits [3:0]"           , isar1,  0, Ext::kDPB, Ext::kDPB2);
  /*
  MERGE_FEATURE_4B("APA bits [7:4]"           , isar1,  4, ...);
  MERGE_FEATURE_4B("API bits [11:8]"          , isar1,  8, ...);
  */
  MERGE_FEATURE_4B("JSCVT bits [15:12]"       , isar1, 12, Ext::kJSCVT);
  MERGE_FEATURE_4B("FCMA bits [19:16]"        , isar1, 16, Ext::kFCMA);
  MERGE_FEATURE_4B("LRCPC bits [23:20]"       , isar1, 20, Ext::kLRCPC, Ext::kLRCPC2, Ext::kLRCPC3);
  /*
  MERGE_FEATURE_4B("GPA bits [27:24]"         , isar1, 24, ...);
  MERGE_FEATURE_4B("GPI bits [31:28]"         , isar1, 28, ...);
  */
  MERGE_FEATURE_4B("FRINTTS bits [35:32]"     , isar1, 32, Ext::kFRINTTS);
  MERGE_FEATURE_4B("SB bits [39:36]"          , isar1, 36, Ext::kSB);
  MERGE_FEATURE_4B("SPECRES bits [43:40]"     , isar1, 40, Ext::kSPECRES, Ext::kSPECRES2);
  MERGE_FEATURE_4B("BF16 bits [47:44]"        , isar1, 44, Ext::kBF16, Ext::kEBF16);
  MERGE_FEATURE_4B("DGH bits [51:48]"         , isar1, 48, Ext::kDGH);
  MERGE_FEATURE_4B("I8MM bits [55:52]"        , isar1, 52, Ext::kI8MM);
  MERGE_FEATURE_4B("XS bits [59:56]"          , isar1, 56, Ext::kXS);
  MERGE_FEATURE_4B("LS64 bits [63:60]"        , isar1, 60, Ext::kLS64, Ext::kLS64_V, Ext::kLS64_ACCDATA);
}

// Detects features based on the content of ID_AA64ISAR2_EL1 register.
[[maybe_unused]]
static inline void detect_aarch64_features_via_cpuid_aa64isar2(CpuInfo& cpu, uint64_t isar2) noexcept {
  MERGE_FEATURE_4B("WFxT bits [3:0]"          , isar2,  0, Ext::kNone, Ext::kWFXT);
  MERGE_FEATURE_4B("RPRES bits [7:4]"         , isar2,  4, Ext::kRPRES);
  /*
  MERGE_FEATURE_4B("GPA3 bits [11:8]"         , isar2,  8, ...);
  MERGE_FEATURE_4B("APA3 bits [15:12]"        , isar2, 12, ...);
  */
  MERGE_FEATURE_4B("MOPS bits [19:16]"        , isar2, 16, Ext::kMOPS);
  MERGE_FEATURE_4B("BC bits [23:20]"          , isar2, 20, Ext::kHBC);
  MERGE_FEATURE_4B("PAC_frac bits [27:24]"    , isar2, 24, Ext::kCONSTPACFIELD);
  MERGE_FEATURE_4B("CLRBHB bits [31:28]"      , isar2, 28, Ext::kCLRBHB);
  MERGE_FEATURE_4B("SYSREG128 bits [35:32]"   , isar2, 32, Ext::kSYSREG128);
  MERGE_FEATURE_4B("SYSINSTR128 bits [39:36]" , isar2, 36, Ext::kSYSINSTR128);
  MERGE_FEATURE_4B("PRFMSLC bits [43:40]"     , isar2, 40, Ext::kPRFMSLC);
  MERGE_FEATURE_4B("RPRFM bits [51:48]"       , isar2, 48, Ext::kRPRFM);
  MERGE_FEATURE_4B("CSSC bits [55:52]"        , isar2, 52, Ext::kCSSC);
  MERGE_FEATURE_4B("LUT bits [59:56]"         , isar2, 56, Ext::kLUT);
  /*
  MERGE_FEATURE_4B("ATS1A bits [63:60]"       , isar2, 60, ...);
  */
}

// TODO: This register is not accessed at the moment.
#if 0
// Detects features based on the content of ID_AA64ISAR3_EL1register.
[[maybe_unused]]
static inline void detect_aarch64_features_via_cpuid_aa64isar3(CpuInfo& cpu, uint64_t isar3) noexcept {
  // ID_AA64ISAR3_EL1
  // ================

  MERGE_FEATURE_4B("CPA bits [3:0]"           , isar3,  0, Ext::kCPA, Ext::kCPA2);
  MERGE_FEATURE_4B("FAMINMAX bits [7:4]"      , isar3,  4, Ext::kFAMINMAX);
  MERGE_FEATURE_4B("TLBIW bits [11:8]"        , isar3,  8, Ext::kTLBIW);
  /*
  MERGE_FEATURE_4B("PACM bits [15:12]"        , isar3, 12, ...);
  */
  MERGE_FEATURE_4B("LSFE bits [19:16]"        , isar3, 16, Ext::kLSFE);
  MERGE_FEATURE_4B("OCCMO bits [23:20]"       , isar3, 20, Ext::kOCCMO);
  MERGE_FEATURE_4B("LSUI bits [27:24]"        , isar3, 24, Ext::kLSUI);
}
#endif

[[maybe_unused]]
static inline void detect_aarch64_features_via_cpuid_aa64mmfr0(CpuInfo& cpu, uint64_t mmfr0) noexcept {
  // ID_AA64MMFR0_EL1
  // ================

  /*
  MERGE_FEATURE_4B("PARange bits [3:0]"       , mmfr0,  0, ...);
  MERGE_FEATURE_4B("ASIDBits bits [7:4]"      , mmfr0,  4, ...);
  MERGE_FEATURE_4B("BigEnd bits [11:8]"       , mmfr0,  8, ...);
  MERGE_FEATURE_4B("SNSMem bits [15:12]"      , mmfr0, 12, ...);
  MERGE_FEATURE_4B("BigEndEL0 bits [19:16]"   , mmfr0, 16, ...);
  MERGE_FEATURE_4B("TGran16 bits [23:20]"     , mmfr0, 20, ...);
  MERGE_FEATURE_4B("TGran64 bits [27:24]"     , mmfr0, 24, ...);
  MERGE_FEATURE_4B("TGran4 bits [31:28]"      , mmfr0, 28, ...);
  MERGE_FEATURE_4B("TGran16_2 bits [35:32]"   , mmfr0, 32, ...);
  MERGE_FEATURE_4B("TGran64_2 bits [39:36]"   , mmfr0, 36, ...);
  MERGE_FEATURE_4B("TGran4_2 bits [43:40]"    , mmfr0, 40, ...);
  MERGE_FEATURE_4B("ExS bits [47:44]"         , mmfr0, 44, ...);
  */
  MERGE_FEATURE_4B("FGT bits [59:56]"         , mmfr0, 56, Ext::kFGT, Ext::kFGT2);
  MERGE_FEATURE_4B("ECV bits [63:60]"         , mmfr0, 60, Ext::kECV);
}

[[maybe_unused]]
static inline void detect_aarch64_features_via_cpuid_aa64mmfr1(CpuInfo& cpu, uint64_t mmfr1) noexcept {
  // ID_AA64MMFR1_EL1
  // ================

  MERGE_FEATURE_4B("HAFDBS bits [3:0]"        , mmfr1,  0, Ext::kHAFDBS, Ext::kNone, Ext::kHAFT, Ext::kHDBSS);
  MERGE_FEATURE_4B("VMIDBits bits [7:4]"      , mmfr1,  4, Ext::kVMID16);
  MERGE_FEATURE_4B("VH bits [11:8]"           , mmfr1,  8, Ext::kVHE);
  MERGE_FEATURE_4B("HPDS bits [15:12]"        , mmfr1, 12, Ext::kHPDS, Ext::kHPDS2);
  MERGE_FEATURE_4B("LO bits [19:16]"          , mmfr1, 16, Ext::kLOR);
  MERGE_FEATURE_4B("PAN bits [23:20]"         , mmfr1, 20, Ext::kPAN, Ext::kPAN2, Ext::kPAN3);
  /*
  MERGE_FEATURE_4B("SpecSEI bits [27:24]"     , mmfr1, 24, ...);
  */
  MERGE_FEATURE_4B("XNX bits [31:28]"         , mmfr1, 28, Ext::kXNX);
  /*
  MERGE_FEATURE_4B("TWED bits [35:32]"        , mmfr1, 32, ...);
  MERGE_FEATURE_4B("ETS bits [39:36]"         , mmfr1, 36, ...);
  */
  MERGE_FEATURE_4B("HCX bits [43:40]"         , mmfr1, 40, Ext::kHCX);
  MERGE_FEATURE_4B("AFP bits [47:44]"         , mmfr1, 44, Ext::kAFP);
  /*
  MERGE_FEATURE_4B("nTLBPA bits [51:48]"      , mmfr1, 48, ...);
  MERGE_FEATURE_4B("TIDCP1 bits [55:52]"      , mmfr1, 52, ...);
  */
  MERGE_FEATURE_4B("CMOW bits [59:56]"        , mmfr1, 56, Ext::kCMOW);
  MERGE_FEATURE_4B("ECBHB bits [63:60]"       , mmfr1, 60, Ext::kECBHB);
}

[[maybe_unused]]
static inline void detect_aarch64_features_via_cpuid_aa64mmfr2(CpuInfo& cpu, uint64_t mmfr2) noexcept {
  // ID_AA64MMFR2_EL1
  // ================

  /*
  MERGE_FEATURE_4B("CnP bits [3:0]"           , mmfr2,  0, ...);
  */
  MERGE_FEATURE_4B("UAO bits [7:4]"           , mmfr2,  4, Ext::kUAO);
  /*
  MERGE_FEATURE_4B("LSM bits [11:8]"          , mmfr2,  8, ...);
  MERGE_FEATURE_4B("IESB bits [15:12]"        , mmfr2, 12, ...);
  */
  MERGE_FEATURE_4B("VARange bits [19:16]"     , mmfr2, 16, Ext::kLVA, Ext::kLVA3);
  MERGE_FEATURE_4B("CCIDX bits [23:20]"       , mmfr2, 20, Ext::kCCIDX);
  MERGE_FEATURE_4B("NV bits [27:24]"          , mmfr2, 24, Ext::kNV, Ext::kNV2);
  /*
  MERGE_FEATURE_4B("ST bits [31:28]"          , mmfr2, 28, ...);
  */
  MERGE_FEATURE_4B("AT bits [35:32]"          , mmfr2, 32, Ext::kLSE2);
  /*
  MERGE_FEATURE_4B("IDS bits [39:36]"         , mmfr2, 36, ...);
  MERGE_FEATURE_4B("FWB bits [43:40]"         , mmfr2, 40, ...);
  MERGE_FEATURE_4B("TTL bits [51:48]"         , mmfr2, 48, ...);
  MERGE_FEATURE_4B("BBM bits [55:52]"         , mmfr2, 52, ...);
  MERGE_FEATURE_4B("EVT bits [59:56]"         , mmfr2, 56, ...);
  MERGE_FEATURE_4B("E0PD bits [63:60]"        , mmfr2, 60, ...);
  */
}

// Detects features based on the content of ID_AA64ZFR0_EL1 (SVE Feature ID register 0).
[[maybe_unused]]
static inline void detect_aarch64_features_via_cpuid_aa64zfr0(CpuInfo& cpu, uint64_t zfr0) noexcept {
  MERGE_FEATURE_4B("SVEver bits [3:0]"        , zfr0,  0, Ext::kSVE2, Ext::kSVE2_1, Ext::kSVE2_2);
  MERGE_FEATURE_4B("AES bits [7:4]"           , zfr0,  4, Ext::kSVE_AES, Ext::kSVE_PMULL128);
  MERGE_FEATURE_4B("EltPerm bits [15:12]"     , zfr0, 12, Ext::kSVE_ELTPERM);
  MERGE_FEATURE_4B("BitPerm bits [19:16]"     , zfr0, 16, Ext::kSVE_BITPERM);
  MERGE_FEATURE_4B("BF16 bits [23:20]"        , zfr0, 20, Ext::kSVE_BF16, Ext::kSVE_EBF16);
  MERGE_FEATURE_4B("B16B16 bits [27:24]"      , zfr0, 24, Ext::kSVE_B16B16);
  MERGE_FEATURE_4B("SHA3 bits [35:32]"        , zfr0, 32, Ext::kSVE_SHA3);
  MERGE_FEATURE_4B("SM4 bits [43:40]"         , zfr0, 40, Ext::kSVE_SM4);
  MERGE_FEATURE_4B("I8MM bits [47:44]"        , zfr0, 44, Ext::kSVE_I8MM);
  MERGE_FEATURE_4B("F32MM bits [55:52]"       , zfr0, 52, Ext::kSVE_F32MM);
  MERGE_FEATURE_4B("F64MM bits [59:56]"       , zfr0, 56, Ext::kSVE_F64MM);
}

[[maybe_unused]]
static inline void detect_aarch64_features_via_cpuid_aa64smfr0(CpuInfo& cpu, uint64_t smfr0) noexcept {
  MERGE_FEATURE_1B("SMOP4 bit [0]"            , smfr0,  0, Ext::kSME_MOP4);
  MERGE_FEATURE_1B("STMOP bit [16]"           , smfr0, 16, Ext::kSME_TMOP);
  MERGE_FEATURE_1B("SFEXPA bit [23]"          , smfr0, 23, Ext::kSSVE_FEXPA);
  MERGE_FEATURE_1B("AES bit [24]"             , smfr0, 24, Ext::kSSVE_AES);
  MERGE_FEATURE_1B("SBitPerm bit [25]"        , smfr0, 25, Ext::kSSVE_BITPERM);
  MERGE_FEATURE_1B("SF8DP2 bit [28]"          , smfr0, 28, Ext::kSSVE_FP8DOT2);
  MERGE_FEATURE_1B("SF8DP4 bit [29]"          , smfr0, 29, Ext::kSSVE_FP8DOT4);
  MERGE_FEATURE_1B("SF8FMA bit [30]"          , smfr0, 30, Ext::kSSVE_FP8FMA);
  MERGE_FEATURE_1B("F32F32 bit [32]"          , smfr0, 32, Ext::kSME_F32F32);
  MERGE_FEATURE_1B("BI32I32 bit [33]"         , smfr0, 33, Ext::kSME_BI32I32);
  MERGE_FEATURE_1B("B16F32 bit [34]"          , smfr0, 34, Ext::kSME_B16F32);
  MERGE_FEATURE_1B("F16F32 bit [35]"          , smfr0, 35, Ext::kSME_F16F32);
  MERGE_FEATURE_4S("I8I32 bits [39:36]"       , smfr0, 36, 0xF, Ext::kSME_I8I32);
  MERGE_FEATURE_1B("F8F32 bit [40]"           , smfr0, 40, Ext::kSME_F8F32);
  MERGE_FEATURE_1B("F8F16 bit [41]"           , smfr0, 41, Ext::kSME_F8F16);
  MERGE_FEATURE_1B("F16F16 bit [42]"          , smfr0, 42, Ext::kSME_F16F16);
  MERGE_FEATURE_1B("B16B16 bit [43]"          , smfr0, 43, Ext::kSME_B16B16);
  MERGE_FEATURE_4S("I16I32 bits [47:44]"      , smfr0, 44, 0x5, Ext::kSME_I16I32);
  MERGE_FEATURE_1B("F64F64 bit [48]"          , smfr0, 48, Ext::kSME_F64F64);
  MERGE_FEATURE_4S("I16I64 bits [55:52]"      , smfr0, 52, 0xF, Ext::kSME_I16I64);
  MERGE_FEATURE_4B("SMEver bits [59:56]"      , smfr0, 56, Ext::kSME2, Ext::kSME2_1, Ext::kSME2_2);
  MERGE_FEATURE_1B("LUTv2 bit [60]"           , smfr0, 60, Ext::kSME_LUTv2);
  MERGE_FEATURE_1B("FA64 bit [63]"            , smfr0, 63, Ext::kSME_FA64);
}

#undef MERGE_FEATURE_4S
#undef MERGE_FEATURE_4B
#undef MERGE_FEATURE_2B
#undef MERGE_FEATURE_1B
#undef MERGE_FEATURE_NA

// CpuInfo - Detect - ARM - CPU Vendor Features
// ============================================

// CPU features detection based on Apple family ID.
enum class AppleFamilyId : uint32_t {
  // Apple design.
  kSWIFT              = 0x1E2D6381u, // Apple A6/A6X  (ARMv7s).
  kCYCLONE            = 0x37A09642u, // Apple A7      (ARMv8.0-A).
  kTYPHOON            = 0x2C91A47Eu, // Apple A8      (ARMv8.0-A).
  kTWISTER            = 0x92FB37C8u, // Apple A9      (ARMv8.0-A).
  kHURRICANE          = 0x67CEEE93u, // Apple A10     (ARMv8.1-A).
  kMONSOON_MISTRAL    = 0xE81E7EF6u, // Apple A11     (ARMv8.2-A).
  kVORTEX_TEMPEST     = 0x07D34B9Fu, // Apple A12     (ARMv8.3-A).
  kLIGHTNING_THUNDER  = 0x462504D2u, // Apple A13     (ARMv8.4-A).
  kFIRESTORM_ICESTORM = 0x1B588BB3u, // Apple A14/M1  (ARMv8.5-A).
  kAVALANCHE_BLIZZARD = 0XDA33D83Du, // Apple A15/M2  (ARMv8.6-A).
  kEVEREST_SAWTOOTH   = 0X8765EDEAu, // Apple A16     (ARMv8.6-A).
  kIBIZA              = 0xFA33415Eu, // Apple M3      (ARMv8.6-A).
  kPALMA              = 0x72015832u, // Apple M3 Max  (ARMv8.6-A).
  kLOBOS              = 0x5F4DEA93u, // Apple M3 Pro  (ARMv8.6-A).
  kCOLL               = 0x2876F5B5u, // Apple A17 Pro (ARMv8.7-A).
  kDONAN              = 0x6F5129ACu, // Apple M4      (ARMv8.7-A).
  kBRAVA              = 0x17D5B93Au, // Apple M4 Max  (ARMv8.7-A).
  kTUPAI              = 0x204526D0u, // Apple A18     .
  kTAHITI             = 0x75D4ACB9u  // Apple A18 Pro .
};

[[maybe_unused]]
static ASMJIT_FAVOR_SIZE bool detect_aarch64_features_via_apple_family_id(CpuInfo& cpu) noexcept {
  using Id = AppleFamilyId;
  CpuFeatures::ARM& features = cpu.features().arm();

  switch (cpu.family_id()) {
    // Apple A7-A9 (ARMv8.0-A).
    case uint32_t(Id::kCYCLONE):
    case uint32_t(Id::kTYPHOON):
    case uint32_t(Id::kTWISTER):
      populate_armv8a_features(features, 0);
      features.add(
        Ext::kPMU,
        Ext::kAES, Ext::kPMULL, Ext::kSHA1, Ext::kSHA256
      );
      return true;

    // Apple A10 (ARMv8.0-A).
    case uint32_t(Id::kHURRICANE):
      populate_armv8a_features(features, 0);
      features.add(
        Ext::kLOR, Ext::kPAN, Ext::kPMU, Ext::kVHE,
        Ext::kAES, Ext::kCRC32, Ext::kPMULL, Ext::kSHA1, Ext::kSHA256, Ext::kRDM
      );
      return true;

    // Apple A11 (ARMv8.2-A).
    case uint32_t(Id::kMONSOON_MISTRAL):
      populate_armv8a_features(features, 2);
      features.add(
        Ext::kPMU,
        Ext::kAES, Ext::kPMULL, Ext::kSHA1, Ext::kSHA256,
        Ext::kFP16, Ext::kFP16CONV
      );
      return true;

    // Apple A12 (ARMv8.3-A).
    case uint32_t(Id::kVORTEX_TEMPEST):
      populate_armv8a_features(features, 3);
      features.add(
        Ext::kPMU,
        Ext::kAES, Ext::kPMULL, Ext::kSHA1, Ext::kSHA256,
        Ext::kFP16, Ext::kFP16CONV
      );
      return true;

    // Apple A13 (ARMv8.4-A).
    case uint32_t(Id::kLIGHTNING_THUNDER):
      populate_armv8a_features(features, 4);
      features.add(
        Ext::kPMU,
        Ext::kFP16, Ext::kFP16CONV, Ext::kFHM,
        Ext::kAES, Ext::kPMULL, Ext::kSHA1, Ext::kSHA256, Ext::kSHA3, Ext::kSHA512
      );
      return true;

    // Apple A14/M1 (ARMv8.5-A).
    case uint32_t(Id::kFIRESTORM_ICESTORM):
      populate_armv8a_features(features, 4);
      features.add(
        Ext::kCSV2, Ext::kCSV3, Ext::kDPB2, Ext::kECV, Ext::kFLAGM2, Ext::kPMU, Ext::kSB, Ext::kSSBS,
        Ext::kFP16, Ext::kFP16CONV, Ext::kFHM,
        Ext::kFRINTTS,
        Ext::kAES, Ext::kPMULL, Ext::kSHA1, Ext::kSHA256, Ext::kSHA3, Ext::kSHA512);
      return true;

    // Apple A15/M2.
    case uint32_t(Id::kAVALANCHE_BLIZZARD):
      populate_armv8a_features(features, 6);
      features.add(
        Ext::kPMU,
        Ext::kFP16, Ext::kFP16CONV, Ext::kFHM,
        Ext::kAES, Ext::kPMULL, Ext::kSHA1, Ext::kSHA256, Ext::kSHA3, Ext::kSHA512);
      return true;

    // Apple A16/M3.
    case uint32_t(Id::kEVEREST_SAWTOOTH):
    case uint32_t(Id::kIBIZA):
    case uint32_t(Id::kPALMA):
    case uint32_t(Id::kLOBOS):
      populate_armv8a_features(features, 6);
      features.add(
        Ext::kHCX, Ext::kPMU,
        Ext::kFP16, Ext::kFP16CONV, Ext::kFHM,
        Ext::kAES, Ext::kPMULL, Ext::kSHA1, Ext::kSHA256, Ext::kSHA3, Ext::kSHA512
      );
      return true;

    // Apple A17/M4.
    case uint32_t(Id::kCOLL):
    case uint32_t(Id::kDONAN):
    case uint32_t(Id::kBRAVA):
      populate_armv8a_features(features, 7);
      features.add(
        Ext::kPMU,
        Ext::kFP16, Ext::kFP16CONV, Ext::kFHM,
        Ext::kAES, Ext::kSHA1, Ext::kSHA3, Ext::kSHA256, Ext::kSHA512,
        Ext::kSME, Ext::kSME2, Ext::kSME_F64F64, Ext::kSME_I16I64);
      return true;

    default:
      return false;
  }
}

// CpuInfo - Detect - ARM - Compile Flags Features
// ===============================================

// Detects ARM version by macros defined at compile time. This means that AsmJit will report features forced at
// compile time that should always be provided by the target CPU. This also means that if we don't provide any
// means to detect CPU features the features reported by AsmJit will at least not report less features than the
// target it was compiled to.

#if ASMJIT_ARCH_ARM == 32
[[maybe_unused]]
static ASMJIT_FAVOR_SIZE void detect_aarch32_features_via_compiler_flags(CpuInfo& cpu) noexcept {
  Support::maybe_unused(cpu);

  // ARM targets have no baseline at the moment.
#if defined(__ARM_ARCH_7A__)
  cpu.add_feature(CpuFeatures::ARM::kARMv7);
#endif

#if defined(__ARM_ARCH_8A__)
  cpu.add_feature(CpuFeatures::ARM::kARMv8a);
#endif

#if defined(__TARGET_ARCH_THUMB)
  cpu.add_feature(CpuFeatures::ARM::kTHUMB);
#if __TARGET_ARCH_THUMB >= 4
  cpu.add_feature(CpuFeatures::ARM::kTHUMBv2);
#endif
#endif

#if defined(__ARM_FEATURE_FMA)
  cpu.add_feature(Ext::kFP);
#endif

#if defined(__ARM_NEON)
  cpu.add_feature(Ext::kASIMD);
#endif

#if defined(__ARM_FEATURE_IDIV) && defined(__TARGET_ARCH_THUMB)
  cpu.add_feature(Ext::kIDIVT);
#endif
#if defined(__ARM_FEATURE_IDIV) && !defined(__TARGET_ARCH_THUMB)
  cpu.add_feature(Ext::kIDIVA);
#endif
}
#endif // ASMJIT_ARCH_ARM == 32

#if ASMJIT_ARCH_ARM == 64
[[maybe_unused]]
static ASMJIT_FAVOR_SIZE void detect_aarch64_features_via_compiler_flags(CpuInfo& cpu) noexcept {
  Support::maybe_unused(cpu);

#if defined(__ARM_ARCH_9_5A__)
  populate_armv9a_features(cpu.features().arm(), 5);
#elif defined(__ARM_ARCH_9_4A__)
  populate_armv9a_features(cpu.features().arm(), 4);
#elif defined(__ARM_ARCH_9_3A__)
  populate_armv9a_features(cpu.features().arm(), 3);
#elif defined(__ARM_ARCH_9_2A__)
  populate_armv9a_features(cpu.features().arm(), 2);
#elif defined(__ARM_ARCH_9_1A__)
  populate_armv9a_features(cpu.features().arm(), 1);
#elif defined(__ARM_ARCH_9A__)
  populate_armv9a_features(cpu.features().arm(), 0);
#elif defined(__ARM_ARCH_8_9A__)
  populate_armv8a_features(cpu.features().arm(), 9);
#elif defined(__ARM_ARCH_8_8A__)
  populate_armv8a_features(cpu.features().arm(), 8);
#elif defined(__ARM_ARCH_8_7A__)
  populate_armv8a_features(cpu.features().arm(), 7);
#elif defined(__ARM_ARCH_8_6A__)
  populate_armv8a_features(cpu.features().arm(), 6);
#elif defined(__ARM_ARCH_8_5A__)
  populate_armv8a_features(cpu.features().arm(), 5);
#elif defined(__ARM_ARCH_8_4A__)
  populate_armv8a_features(cpu.features().arm(), 4);
#elif defined(__ARM_ARCH_8_3A__)
  populate_armv8a_features(cpu.features().arm(), 3);
#elif defined(__ARM_ARCH_8_2A__)
  populate_armv8a_features(cpu.features().arm(), 2);
#elif defined(__ARM_ARCH_8_1A__)
  populate_armv8a_features(cpu.features().arm(), 1);
#else
  populate_armv8a_features(cpu.features().arm(), 0);
#endif

#if defined(__ARM_FEATURE_AES)
  cpu.add_feature(Ext::kAES);
#endif

#if defined(__ARM_FEATURE_BF16_SCALAR_ARITHMETIC) && defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC)
  cpu.add_feature(Ext::kBF16);
#endif

#if defined(__ARM_FEATURE_CRC32)
  cpu.add_feature(Ext::kCRC32);
#endif

#if defined(__ARM_FEATURE_CRYPTO)
  cpu.add_feature(Ext::kAES, Ext::kSHA1, Ext::kSHA256);
#endif

#if defined(__ARM_FEATURE_DOTPROD)
  cpu.add_feature(Ext::kDOTPROD);
#endif

#if defined(__ARM_FEATURE_FP16FML) || defined(__ARM_FEATURE_FP16_FML)
  cpu.add_feature(Ext::kFHM);
#endif

#if defined(__ARM_FEATURE_FP16_SCALAR_ARITHMETIC)
  cpu.add_feature(Ext::kFP16);
#endif

#if defined(__ARM_FEATURE_FRINT)
  cpu.add_feature(Ext::kFRINTTS);
#endif

#if defined(__ARM_FEATURE_JCVT)
  cpu.add_feature(Ext::kJSCVT);
#endif

#if defined(__ARM_FEATURE_MATMUL_INT8)
  cpu.add_feature(Ext::kI8MM);
#endif

#if defined(__ARM_FEATURE_ATOMICS)
  cpu.add_feature(Ext::kLSE);
#endif

#if defined(__ARM_FEATURE_MEMORY_TAGGING)
  cpu.add_feature(Ext::kMTE);
#endif

#if defined(__ARM_FEATURE_QRDMX)
  cpu.add_feature(Ext::kRDM);
#endif

#if defined(__ARM_FEATURE_RNG)
  cpu.add_feature(Ext::kRNG);
#endif

#if defined(__ARM_FEATURE_SHA2)
  cpu.add_feature(Ext::kSHA256);
#endif

#if defined(__ARM_FEATURE_SHA3)
  cpu.add_feature(Ext::kSHA3);
#endif

#if defined(__ARM_FEATURE_SHA512)
  cpu.add_feature(Ext::kSHA512);
#endif

#if defined(__ARM_FEATURE_SM3)
  cpu.add_feature(Ext::kSM3);
#endif

#if defined(__ARM_FEATURE_SM4)
  cpu.add_feature(Ext::kSM4);
#endif

#if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_SVE_VECTOR_OPERATORS)
  cpu.add_feature(Ext::kSVE);
#endif

#if defined(__ARM_FEATURE_SVE_MATMUL_INT8)
  cpu.add_feature(Ext::kSVE_I8MM);
#endif

#if defined(__ARM_FEATURE_SVE_MATMUL_FP32)
  cpu.add_feature(Ext::kSVE_F32MM);
#endif

#if defined(__ARM_FEATURE_SVE_MATMUL_FP64)
  cpu.add_feature(Ext::kSVE_F64MM);
#endif

#if defined(__ARM_FEATURE_SVE2)
  cpu.add_feature(Ext::kSVE2);
#endif

#if defined(__ARM_FEATURE_SVE2_AES)
  cpu.add_feature(Ext::kSVE_AES);
#endif

#if defined(__ARM_FEATURE_SVE2_BITPERM)
  cpu.add_feature(Ext::kSVE_BITPERM);
#endif

#if defined(__ARM_FEATURE_SVE2_SHA3)
  cpu.add_feature(Ext::kSVE_SHA3);
#endif

#if defined(__ARM_FEATURE_SVE2_SM4)
  cpu.add_feature(Ext::kSVE_SM4);
#endif

#if defined(__ARM_FEATURE_TME)
  cpu.add_feature(Ext::kTME);
#endif
}
#endif // ASMJIT_ARCH_ARM == 64

[[maybe_unused]]
static ASMJIT_FAVOR_SIZE void detect_arm_features_via_compiler_flags(CpuInfo& cpu) noexcept {
#if ASMJIT_ARCH_ARM == 32
  detect_aarch32_features_via_compiler_flags(cpu);
#else
  detect_aarch64_features_via_compiler_flags(cpu);
#endif // ASMJIT_ARCH_ARM
}

// CpuInfo - Detect - ARM - Post Processing ARM Features
// =====================================================

// Postprocesses AArch32 features.
[[maybe_unused]]
static ASMJIT_FAVOR_SIZE void post_process_aarch32_features(CpuFeatures::ARM& features) noexcept {
  Support::maybe_unused(features);
}

// Postprocesses AArch64 features.
//
// The only reason to use this function is to deduce some flags from others.
[[maybe_unused]]
static ASMJIT_FAVOR_SIZE void post_process_aarch64_features(CpuFeatures::ARM& features) noexcept {
  if (features.has_fp16()) {
    features.add(Ext::kFP16CONV);
  }

  if (features.has_mte3()) {
    features.add(Ext::kMTE2);
  }

  if (features.has_mte2()) {
    features.add(Ext::kMTE);
  }

  if (features.has_ssbs2()) {
    features.add(Ext::kSSBS);
  }
}

[[maybe_unused]]
static ASMJIT_FAVOR_SIZE void post_process_arm_cpu_info(CpuInfo& cpu) noexcept {
#if ASMJIT_ARCH_ARM == 32
  post_process_aarch32_features(cpu.features().arm());
#else
  post_process_aarch64_features(cpu.features().arm());
#endif // ASMJIT_ARCH_ARM
}

// CpuInfo - Detect - ARM - Detect by Reading CPUID Registers
// ==========================================================

// Support CPUID-based detection on AArch64.
#if defined(ASMJIT_ARM_DETECT_VIA_CPUID)

// Since the register ID is encoded with the instruction we have to create a function for each register ID to read.
#define ASMJIT_AARCH64_DEFINE_CPUID_READ_FN(func, reg_id)  \
[[maybe_unused]]                                           \
static inline uint64_t func() noexcept {                   \
  uint64_t output;                                         \
  __asm__ __volatile__("mrs %0, " #reg_id : "=r"(output)); \
  return output;                                           \
}

// NOTE: Older tools don't know the IDs. For example Ubuntu on RPI (GCC 9) won't compile ID_AA64ISAR2_EL1 in 2023.
ASMJIT_AARCH64_DEFINE_CPUID_READ_FN(aarch64_read_pfr0, ID_AA64PFR0_EL1)
ASMJIT_AARCH64_DEFINE_CPUID_READ_FN(aarch64_read_pfr1, ID_AA64PFR1_EL1)
ASMJIT_AARCH64_DEFINE_CPUID_READ_FN(aarch64_read_isar0, ID_AA64ISAR0_EL1)
ASMJIT_AARCH64_DEFINE_CPUID_READ_FN(aarch64_read_isar1, ID_AA64ISAR1_EL1)
ASMJIT_AARCH64_DEFINE_CPUID_READ_FN(aarch64_read_isar2, S3_0_C0_C6_2)     // ID_AA64ISAR2_EL1
ASMJIT_AARCH64_DEFINE_CPUID_READ_FN(aarch64_read_zfr0, S3_0_C0_C4_4)      // ID_AA64ZFR0_EL1

#undef ASMJIT_AARCH64_DEFINE_CPUID_READ_FN

// Detects AArch64 features by reading CPUID bits directly from CPUID registers. This is the most reliable method
// as the OS doesn't have to know all supported extensions this way (if there is something missing in HWCAPS then
// there is no way to detect such feature without reading CPUID bits).
//
// This function uses MSR instructions, which means that it reads registers that cannot be read in user-mode. The
// OS typically implements this feature by handling SIGILL internally and providing a filtered content of these
// registers back to the user - at least this is what Linux documentation states - everything implementation
// dependent is zeroed, only the bits that are used for CPU feature identification would be present.
//
// References:
//   - https://docs.kernel.org/arch/arm64/cpu-feature-registers.html
[[maybe_unused]]
static ASMJIT_FAVOR_SIZE void detect_aarch64_features_via_cpuid(CpuInfo& cpu) noexcept {
  populate_base_arm_features(cpu);

  detect_aarch64_features_via_cpuid_aa64pfr0_aa64pfr1(cpu, aarch64_read_pfr0(), aarch64_read_pfr1());
  detect_aarch64_features_via_cpuid_aa64isar0_aa64isar1(cpu, aarch64_read_isar0(), aarch64_read_isar1());

  // TODO: Fix this on FreeBSD - I don't know what kernel version allows to access the registers below...

#if defined(__linux__)
  UNameKernelVersion kVer = get_kernel_version_via_uname();

  // Introduced in Linux 4.19 by "arm64: add ID_AA64ISAR2_EL1 sys register"), so we want at least 4.20.
  if (kVer.at_least(4, 20)) {
    detect_aarch64_features_via_cpuid_aa64isar2(cpu, aarch64_read_isar2());
  }

  // Introduced in Linux 5.10 by "arm64: Expose SVE2 features for userspace", so we want at least 5.11.
  if (kVer.at_least(5, 11) && cpu.features().arm().has_any(Ext::kSVE, Ext::kSME)) {
    // Only read CPU_ID_AA64ZFR0 when either SVE or SME is available.
    detect_aarch64_features_via_cpuid_aa64zfr0(cpu, aarch64_read_zfr0());
  }
#endif
}

#endif // ASMJIT_ARM_DETECT_VIA_CPUID

// CpuInfo - Detect - ARM - Detect by Windows API
// ==============================================

#if defined(_WIN32)
struct WinPFPMapping {
  uint8_t feature_id;
  uint8_t pfp_feature_id;
};

static ASMJIT_FAVOR_SIZE void detect_pfp_features(CpuInfo& cpu, const WinPFPMapping* mapping, size_t size) noexcept {
  for (size_t i = 0; i < size; i++) {
    if (::IsProcessorFeaturePresent(mapping[i].pfp_feature_id)) {
      cpu.add_feature(mapping[i].feature_id);
    }
  }
}

//! Detect ARM CPU features on Windows.
//!
//! The detection is based on `IsProcessorFeaturePresent()` API call.
static ASMJIT_FAVOR_SIZE void detect_arm_cpu(CpuInfo& cpu) noexcept {
  cpu._was_detected = true;
  populate_base_arm_features(cpu);

  CpuFeatures::ARM& features = cpu.features().arm();

  // Win32 for ARM requires ARMv7 with DSP extensions, VFPv3 (FP), and uses THUMBv2 by default.
#if ASMJIT_ARCH_ARM == 32
  features.add(Ext::kTHUMB);
  features.add(Ext::kTHUMBv2);
  features.add(Ext::kARMv6);
  features.add(Ext::kARMv7);
  features.add(Ext::kEDSP);
#endif

  // Windows for ARM requires FP and ASIMD.
  features.add(Ext::kFP);
  features.add(Ext::kASIMD);

  // Detect additional CPU features by calling `IsProcessorFeaturePresent()`.
  static const WinPFPMapping mapping[] = {
#if ASMJIT_ARCH_ARM == 32
    { uint8_t(Ext::kVFP_D32)  , 18 }, // PF_ARM_VFP_32_REGISTERS_AVAILABLE
    { uint8_t(Ext::kIDIVT)    , 24 }, // PF_ARM_DIVIDE_INSTRUCTION_AVAILABLE
    { uint8_t(Ext::kFMAC)     , 27 }, // PF_ARM_FMAC_INSTRUCTIONS_AVAILABLE
    { uint8_t(Ext::kARMv8a)   , 29 }, // PF_ARM_V8_INSTRUCTIONS_AVAILABLE
#endif
    { uint8_t(Ext::kAES)      , 30 }, // PF_ARM_V8_CRYPTO_INSTRUCTIONS_AVAILABLE
    { uint8_t(Ext::kCRC32)    , 31 }, // PF_ARM_V8_CRC32_INSTRUCTIONS_AVAILABLE
    { uint8_t(Ext::kLSE)      , 34 }, // PF_ARM_V81_ATOMIC_INSTRUCTIONS_AVAILABLE
    { uint8_t(Ext::kDOTPROD)  , 43 }, // PF_ARM_V82_DP_INSTRUCTIONS_AVAILABLE
    { uint8_t(Ext::kJSCVT)    , 44 }, // PF_ARM_V83_JSCVT_INSTRUCTIONS_AVAILABLE
    { uint8_t(Ext::kLRCPC)    , 45 }  // PF_ARM_V83_LRCPC_INSTRUCTIONS_AVAILABLE
  };
  detect_pfp_features(cpu, mapping, ASMJIT_ARRAY_SIZE(mapping));

  // Windows can only report ARMv8A at the moment.
  if (features.has_armv8a()) {
    populate_armv8a_features(cpu.features().arm(), 0);
  }

  // Windows provides several instructions under a single flag:
  if (features.has_aes()) {
    features.add(Ext::kPMULL, Ext::kSHA1, Ext::kSHA256);
  }

  post_process_arm_cpu_info(cpu);
}

// CpuInfo - Detect - ARM - Detect by Reading HWCAPS
// =================================================

#elif defined(ASMJIT_ARM_DETECT_VIA_HWCAPS)

#ifndef AT_HWCAP
  #define AT_HWCAP 16
#endif // !AT_HWCAP

#ifndef AT_HWCAP2
  #define AT_HWCAP2 26
#endif // !AT_HWCAP2

#if defined(__linux__)
static void get_aux_values(unsigned long* vals, const unsigned long* tags, size_t count) noexcept {
  for (size_t i = 0; i < count; i++) {
    vals[i] = getauxval(tags[i]);
  }
}
#elif defined(__FreeBSD__)
static void get_aux_values(unsigned long* vals, const unsigned long* tags, size_t count) noexcept {
  for (size_t i = 0; i < count; i++) {
    unsigned long result = 0;
    if (elf_aux_info(int(tags[i]), &result, int(sizeof(unsigned long))) != 0)
      result = 0;
    vals[i] = result;
  }
}
#else
#error "[asmjit] get_aux_values() - Unsupported OS."
#endif

static const unsigned long hw_cap_tags_table[2] = { AT_HWCAP, AT_HWCAP2 };

struct HWCapMapping32 { uint8_t feature_id[32]; };
struct HWCapMapping64 { uint8_t feature_id[64]; };

template<typename Mask, typename Map>
static ASMJIT_FAVOR_SIZE void merge_hw_caps(CpuInfo& cpu, const Mask& mask, const Map& map) noexcept {
  static_assert(sizeof(Mask) * 8u == sizeof(Map));

  Support::BitWordIterator<Mask> it(mask);
  while (it.has_next()) {
    uint32_t feature_id = map.feature_id[it.next()];
    cpu.features().add(feature_id);
  }
  cpu.features().remove(0xFFu);
}

#if ASMJIT_ARCH_ARM == 32

// Reference:
//   - https://github.com/torvalds/linux/blob/master/arch/arm/include/uapi/asm/hwcap.h
static constexpr HWCapMapping32 hw_cap1_mapping_table = {{
  uint8_t(0xFF)               , // [ 0]
  uint8_t(0xFF)               , // [ 1]
  uint8_t(0xFF)               , // [ 2]
  uint8_t(0xFF)               , // [ 3]
  uint8_t(0xFF)               , // [ 4]
  uint8_t(0xFF)               , // [ 5]
  uint8_t(0xFF)               , // [ 6]
  uint8_t(Ext::kEDSP)         , // [ 7] HWCAP_EDSP
  uint8_t(0xFF)               , // [ 8]
  uint8_t(0xFF)               , // [ 9]
  uint8_t(0xFF)               , // [10]
  uint8_t(0xFF)               , // [11]
  uint8_t(Ext::kASIMD)        , // [12] HWCAP_NEON
  uint8_t(Ext::kFP)           , // [13] HWCAP_VFPv3
  uint8_t(0xFF)               , // [14]
  uint8_t(0xFF)               , // [15]
  uint8_t(Ext::kFMAC)         , // [16] HWCAP_VFPv4
  uint8_t(Ext::kIDIVA)        , // [17] HWCAP_IDIVA
  uint8_t(Ext::kIDIVT)        , // [18] HWCAP_IDIVT
  uint8_t(Ext::kVFP_D32)      , // [19] HWCAP_VFPD32
  uint8_t(0xFF)               , // [20]
  uint8_t(0xFF)               , // [21]
  uint8_t(Ext::kFP16CONV)     , // [22] HWCAP_FPHP
  uint8_t(Ext::kFP16)         , // [23] HWCAP_ASIMDHP
  uint8_t(Ext::kDOTPROD)      , // [24] HWCAP_ASIMDDP
  uint8_t(Ext::kFHM)          , // [25] HWCAP_ASIMDFHM
  uint8_t(Ext::kBF16)         , // [26] HWCAP_ASIMDBF16
  uint8_t(Ext::kI8MM)         , // [27] HWCAP_I8MM
  uint8_t(0xFF)               , // [28]
  uint8_t(0xFF)               , // [29]
  uint8_t(0xFF)               , // [30]
  uint8_t(0xFF)                 // [31]
}};

static constexpr HWCapMapping32 hw_cap2_mapping_table = {{
  uint8_t(Ext::kAES)          , // [ 0] HWCAP2_AES
  uint8_t(Ext::kPMULL)        , // [ 1] HWCAP2_PMULL
  uint8_t(Ext::kSHA1)         , // [ 2] HWCAP2_SHA1
  uint8_t(Ext::kSHA256)       , // [ 3] HWCAP2_SHA2
  uint8_t(Ext::kCRC32)        , // [ 4] HWCAP2_CRC32
  uint8_t(Ext::kSB)           , // [ 5] HWCAP2_SB
  uint8_t(Ext::kSSBS)         , // [ 6] HWCAP2_SSBS
  uint8_t(0xFF)               , // [ 7]
  uint8_t(0xFF)               , // [ 8]
  uint8_t(0xFF)               , // [ 9]
  uint8_t(0xFF)               , // [10]
  uint8_t(0xFF)               , // [11]
  uint8_t(0xFF)               , // [12]
  uint8_t(0xFF)               , // [13]
  uint8_t(0xFF)               , // [14]
  uint8_t(0xFF)               , // [15]
  uint8_t(0xFF)               , // [16]
  uint8_t(0xFF)               , // [17]
  uint8_t(0xFF)               , // [18]
  uint8_t(0xFF)               , // [19]
  uint8_t(0xFF)               , // [20]
  uint8_t(0xFF)               , // [21]
  uint8_t(0xFF)               , // [22]
  uint8_t(0xFF)               , // [23]
  uint8_t(0xFF)               , // [24]
  uint8_t(0xFF)               , // [25]
  uint8_t(0xFF)               , // [26]
  uint8_t(0xFF)               , // [27]
  uint8_t(0xFF)               , // [28]
  uint8_t(0xFF)               , // [29]
  uint8_t(0xFF)               , // [30]
  uint8_t(0xFF)                 // [31]
}};

static ASMJIT_FAVOR_SIZE void detect_arm_cpu(CpuInfo& cpu) noexcept {
  cpu._was_detected = true;
  populate_base_arm_features(cpu);

  unsigned long hw_cap_masks[2] {};
  get_aux_values(hw_cap_masks, hw_cap_tags_table, 2u);

  merge_hw_caps(cpu, hw_cap_masks[0], hw_cap1_mapping_table);
  merge_hw_caps(cpu, hw_cap_masks[1], hw_cap2_mapping_table);

  CpuFeatures::ARM& features = cpu.features().arm();

  // ARMv7 provides FP|ASIMD.
  if (features.has_fp() || features.has_asimd()) {
    features.add(CpuFeatures::ARM::kARMv7);
  }

  // ARMv8 provives AES, CRC32, PMULL, SHA1, and SHA256.
  if (features.has_aes() || features.has_crc32() || features.has_pmull() || features.has_sha1() || features.has_sha256()) {
    features.add(CpuFeatures::ARM::kARMv8a);
  }

  post_process_arm_cpu_info(cpu);
}

#else

// Reference:
//   - https://docs.kernel.org/arch/arm64/elf_hwcaps.html
//   - https://github.com/torvalds/linux/blob/master/arch/arm64/include/uapi/asm/hwcap.h
static constexpr HWCapMapping64 hw_cap1_mapping_table = {{
  uint8_t(Ext::kFP)           , // [ 0] HWCAP_FP
  uint8_t(Ext::kASIMD)        , // [ 1] HWCAP_ASIMD
  uint8_t(0xFF)               , // [ 2] HWCAP_EVTSTRM
  uint8_t(Ext::kAES)          , // [ 3] HWCAP_AES
  uint8_t(Ext::kPMULL)        , // [ 4] HWCAP_PMULL
  uint8_t(Ext::kSHA1)         , // [ 5] HWCAP_SHA1
  uint8_t(Ext::kSHA256)       , // [ 6] HWCAP_SHA2
  uint8_t(Ext::kCRC32)        , // [ 7] HWCAP_CRC32
  uint8_t(Ext::kLSE)          , // [ 8] HWCAP_ATOMICS
  uint8_t(Ext::kFP16CONV)     , // [ 9] HWCAP_FPHP
  uint8_t(Ext::kFP16)         , // [10] HWCAP_ASIMDHP
  uint8_t(Ext::kCPUID)        , // [11] HWCAP_CPUID
  uint8_t(Ext::kRDM)          , // [12] HWCAP_ASIMDRDM
  uint8_t(Ext::kJSCVT)        , // [13] HWCAP_JSCVT
  uint8_t(Ext::kFCMA)         , // [14] HWCAP_FCMA
  uint8_t(Ext::kLRCPC)        , // [15] HWCAP_LRCPC
  uint8_t(Ext::kDPB)          , // [16] HWCAP_DCPOP
  uint8_t(Ext::kSHA3)         , // [17] HWCAP_SHA3
  uint8_t(Ext::kSM3)          , // [18] HWCAP_SM3
  uint8_t(Ext::kSM4)          , // [19] HWCAP_SM4
  uint8_t(Ext::kDOTPROD)      , // [20] HWCAP_ASIMDDP
  uint8_t(Ext::kSHA512)       , // [21] HWCAP_SHA512
  uint8_t(Ext::kSVE)          , // [22] HWCAP_SVE
  uint8_t(Ext::kFHM)          , // [23] HWCAP_ASIMDFHM
  uint8_t(Ext::kDIT)          , // [24] HWCAP_DIT
  uint8_t(Ext::kLSE2)         , // [25] HWCAP_USCAT
  uint8_t(Ext::kLRCPC2)       , // [26] HWCAP_ILRCPC
  uint8_t(Ext::kFLAGM)        , // [27] HWCAP_FLAGM
  uint8_t(Ext::kSSBS)         , // [28] HWCAP_SSBS
  uint8_t(Ext::kSB)           , // [29] HWCAP_SB
  uint8_t(0xFF)               , // [30] HWCAP_PACA
  uint8_t(0xFF)               , // [31] HWCAP_PACG
  uint8_t(Ext::kGCS)          , // [32] HWCAP_GCS
  uint8_t(Ext::kCMPBR)        , // [33] HWCAP_CMPBR
  uint8_t(Ext::kFPRCVT)       , // [34] HWCAP_FPRCVT
  uint8_t(Ext::kF8F32MM)      , // [35] HWCAP_F8MM8
  uint8_t(Ext::kF8F16MM)      , // [36] HWCAP_F8MM4
  uint8_t(Ext::kSVE_F16MM)    , // [37] HWCAP_SVE_F16MM
  uint8_t(Ext::kSVE_ELTPERM)  , // [38] HWCAP_SVE_ELTPERM
  uint8_t(Ext::kSVE_AES2)     , // [39] HWCAP_SVE_AES2
  uint8_t(Ext::kSVE_BFSCALE)  , // [40] HWCAP_SVE_BFSCALE
  uint8_t(Ext::kSVE2_2)       , // [41] HWCAP_SVE2P2
  uint8_t(Ext::kSME2_2)       , // [42] HWCAP_SME2P2
  uint8_t(Ext::kSSVE_BITPERM) , // [43] HWCAP_SME_SBITPERM
  uint8_t(Ext::kSME_AES)      , // [44] HWCAP_SME_AES
  uint8_t(Ext::kSSVE_FEXPA)   , // [45] HWCAP_SME_SFEXPA
  uint8_t(Ext::kSME_TMOP)     , // [46] HWCAP_SME_STMOP
  uint8_t(Ext::kSME_MOP4)     , // [47] HWCAP_SME_SMOP4
  uint8_t(0xFF)               , // [48]
  uint8_t(0xFF)               , // [49]
  uint8_t(0xFF)               , // [50]
  uint8_t(0xFF)               , // [51]
  uint8_t(0xFF)               , // [52]
  uint8_t(0xFF)               , // [53]
  uint8_t(0xFF)               , // [54]
  uint8_t(0xFF)               , // [55]
  uint8_t(0xFF)               , // [56]
  uint8_t(0xFF)               , // [57]
  uint8_t(0xFF)               , // [58]
  uint8_t(0xFF)               , // [59]
  uint8_t(0xFF)               , // [60]
  uint8_t(0xFF)               , // [61]
  uint8_t(0xFF)               , // [62]
  uint8_t(0xFF)                 // [63]
}};

static constexpr HWCapMapping64 hw_cap2_mapping_table = {{
  uint8_t(Ext::kDPB2)         , // [ 0] HWCAP2_DCPODP
  uint8_t(Ext::kSVE2)         , // [ 1] HWCAP2_SVE2
  uint8_t(Ext::kSVE_AES)      , // [ 2] HWCAP2_SVEAES
  uint8_t(Ext::kSVE_PMULL128) , // [ 3] HWCAP2_SVEPMULL
  uint8_t(Ext::kSVE_BITPERM)  , // [ 4] HWCAP2_SVEBITPERM
  uint8_t(Ext::kSVE_SHA3)     , // [ 5] HWCAP2_SVESHA3
  uint8_t(Ext::kSVE_SM4)      , // [ 6] HWCAP2_SVESM4
  uint8_t(Ext::kFLAGM2)       , // [ 7] HWCAP2_FLAGM2
  uint8_t(Ext::kFRINTTS)      , // [ 8] HWCAP2_FRINT
  uint8_t(Ext::kSVE_I8MM)     , // [ 9] HWCAP2_SVEI8MM
  uint8_t(Ext::kSVE_F32MM)    , // [10] HWCAP2_SVEF32MM
  uint8_t(Ext::kSVE_F64MM)    , // [11] HWCAP2_SVEF64MM
  uint8_t(Ext::kSVE_BF16)     , // [12] HWCAP2_SVEBF16
  uint8_t(Ext::kI8MM)         , // [13] HWCAP2_I8MM
  uint8_t(Ext::kBF16)         , // [14] HWCAP2_BF16
  uint8_t(Ext::kDGH)          , // [15] HWCAP2_DGH
  uint8_t(Ext::kRNG)          , // [16] HWCAP2_RNG
  uint8_t(Ext::kBTI)          , // [17] HWCAP2_BTI
  uint8_t(Ext::kMTE)          , // [18] HWCAP2_MTE
  uint8_t(Ext::kECV)          , // [19] HWCAP2_ECV
  uint8_t(Ext::kAFP)          , // [20] HWCAP2_AFP
  uint8_t(Ext::kRPRES)        , // [21] HWCAP2_RPRES
  uint8_t(Ext::kMTE3)         , // [22] HWCAP2_MTE3
  uint8_t(Ext::kSME)          , // [23] HWCAP2_SME
  uint8_t(Ext::kSME_I16I64)   , // [24] HWCAP2_SME_I16I64
  uint8_t(Ext::kSME_F64F64)   , // [25] HWCAP2_SME_F64F64
  uint8_t(Ext::kSME_I8I32)    , // [26] HWCAP2_SME_I8I32
  uint8_t(Ext::kSME_F16F32)   , // [27] HWCAP2_SME_F16F32
  uint8_t(Ext::kSME_B16F32)   , // [28] HWCAP2_SME_B16F32
  uint8_t(Ext::kSME_F32F32)   , // [29] HWCAP2_SME_F32F32
  uint8_t(Ext::kSME_FA64)     , // [30] HWCAP2_SME_FA64
  uint8_t(Ext::kWFXT)         , // [31] HWCAP2_WFXT
  uint8_t(Ext::kEBF16)        , // [32] HWCAP2_EBF16
  uint8_t(Ext::kSVE_EBF16)    , // [33] HWCAP2_SVE_EBF16
  uint8_t(Ext::kCSSC)         , // [34] HWCAP2_CSSC
  uint8_t(Ext::kRPRFM)        , // [35] HWCAP2_RPRFM
  uint8_t(Ext::kSVE2_1)       , // [36] HWCAP2_SVE2P1
  uint8_t(Ext::kSME2)         , // [37] HWCAP2_SME2
  uint8_t(Ext::kSME2_1)       , // [38] HWCAP2_SME2P1
  uint8_t(Ext::kSME_I16I32)   , // [39] HWCAP2_SME_I16I32
  uint8_t(Ext::kSME_BI32I32)  , // [40] HWCAP2_SME_BI32I32
  uint8_t(Ext::kSME_B16B16)   , // [41] HWCAP2_SME_B16B16
  uint8_t(Ext::kSME_F16F16)   , // [42] HWCAP2_SME_F16F16
  uint8_t(Ext::kMOPS)         , // [43] HWCAP2_MOPS
  uint8_t(Ext::kHBC)          , // [44] HWCAP2_HBC
  uint8_t(Ext::kSVE_B16B16)   , // [45] HWCAP2_SVE_B16B16
  uint8_t(Ext::kLRCPC3)       , // [46] HWCAP2_LRCPC3
  uint8_t(Ext::kLSE128)       , // [47] HWCAP2_LSE128
  uint8_t(Ext::kFPMR)         , // [48] HWCAP2_FPMR
  uint8_t(Ext::kLUT)          , // [49] HWCAP2_LUT
  uint8_t(Ext::kFAMINMAX)     , // [50] HWCAP2_FAMINMAX
  uint8_t(Ext::kFP8)          , // [51] HWCAP2_F8CVT
  uint8_t(Ext::kFP8FMA)       , // [52] HWCAP2_F8FMA
  uint8_t(Ext::kFP8DOT4)      , // [53] HWCAP2_F8DP4
  uint8_t(Ext::kFP8DOT2)      , // [54] HWCAP2_F8DP2
  uint8_t(Ext::kF8E4M3)       , // [55] HWCAP2_F8E4M3
  uint8_t(Ext::kF8E5M2)       , // [56] HWCAP2_F8E5M2
  uint8_t(Ext::kSME_LUTv2)    , // [57] HWCAP2_SME_LUTV2
  uint8_t(Ext::kSME_F8F16)    , // [58] HWCAP2_SME_F8F16
  uint8_t(Ext::kSME_F8F32)    , // [59] HWCAP2_SME_F8F32
  uint8_t(Ext::kSSVE_FP8FMA)  , // [60] HWCAP2_SME_SF8FMA
  uint8_t(Ext::kSSVE_FP8DOT4) , // [61] HWCAP2_SME_SF8DP4
  uint8_t(Ext::kSSVE_FP8DOT2) , // [62] HWCAP2_SME_SF8DP2
  uint8_t(0xFF)                 // [63] HWCAP2_POE
}};

static ASMJIT_FAVOR_SIZE void detect_arm_cpu(CpuInfo& cpu) noexcept {
  cpu._was_detected = true;
  populate_base_arm_features(cpu);

  unsigned long hw_cap_masks[2] {};
  get_aux_values(hw_cap_masks, hw_cap_tags_table, 2u);

  merge_hw_caps(cpu, hw_cap_masks[0], hw_cap1_mapping_table);
  merge_hw_caps(cpu, hw_cap_masks[1], hw_cap2_mapping_table);

#if defined(ASMJIT_ARM_DETECT_VIA_CPUID)
  if (cpu.features().arm().has_cpuid()) {
    detect_aarch64_features_via_cpuid(cpu);
    return;
  }
#endif // ASMJIT_ARM_DETECT_VIA_CPUID

  post_process_arm_cpu_info(cpu);
}

#endif // ASMJIT_ARCH_ARM

// CpuInfo - Detect - ARM - Detect by NetBSD API That Reads CPUID
// ==============================================================

#elif defined(__NetBSD__) && ASMJIT_ARCH_ARM >= 64

//! Position of AArch64 registers in a`aarch64_sysctl_cpu_id` struct, which is filled by sysctl().
struct NetBSDAArch64Regs {
  enum ID : uint32_t {
    k64_MIDR      = 0,    //!< Main ID Register.
    k64_REVIDR    = 8,    //!< Revision ID Register.
    k64_MPIDR     = 16,   //!< Multiprocessor Affinity Register.
    k64_AA64DFR0  = 24,   //!< A64 Debug Feature Register 0.
    k64_AA64DFR1  = 32,   //!< A64 Debug Feature Register 1.
    k64_AA64ISAR0 = 40,   //!< A64 Instruction Set Attribute Register 0.
    k64_AA64ISAR1 = 48,   //!< A64 Instruction Set Attribute Register 1.
    k64_AA64MMFR0 = 56,   //!< A64 Memory Model Feature Register 0.
    k64_AA64MMFR1 = 64,   //!< A64 Memory Model Feature Register 1.
    k64_AA64MMFR2 = 72,   //!< A64 Memory Model Feature Register 2.
    k64_AA64PFR0  = 80,   //!< A64 Processor Feature Register 0.
    k64_AA64PFR1  = 88,   //!< A64 Processor Feature Register 1.
    k64_AA64ZFR0  = 96,   //!< A64 SVE Feature ID Register 0.
    k32_MVFR0     = 104,  //!< Media and VFP Feature Register 0.
    k32_MVFR1     = 108,  //!< Media and VFP Feature Register 1.
    k32_MVFR2     = 112,  //!< Media and VFP Feature Register 2.
    k32_PAD       = 116,  //!< Padding (not used).
    k64_CLIDR     = 120,  //!< Cache Level ID Register.
    k64_CTR       = 128   //!< Cache Type Register.
  };

  enum Limits : uint32_t {
    kBufferSize   = 136
  };

  uint64_t data[kBufferSize / 8u];

  ASMJIT_INLINE_NODEBUG uint64_t r64(uint32_t index) const noexcept {
    ASMJIT_ASSERT(index % 8u == 0u);
    return data[index / 8u];
  }

  ASMJIT_INLINE_NODEBUG uint32_t r32(uint32_t index) const noexcept {
    ASMJIT_ASSERT(index % 4u == 0u);
    uint32_t shift = (index % 8) * 8;
    return uint32_t((r64(index) >> shift) & 0xFFFFFFFFu);
  }
};

static ASMJIT_FAVOR_SIZE void detect_arm_cpu(CpuInfo& cpu) noexcept {
  using Regs = NetBSDAArch64Regs;

  populate_base_arm_features(cpu);

  Regs regs {};
  size_t len = sizeof(regs);
  const char sysctl_cpu_path[] = "machdep.cpu0.cpu_id";

  if (sysctlbyname(sysctl_cpu_path, &regs, &len, nullptr, 0) == 0) {
    detect_aarch64_features_via_cpuid_aa64pfr0_aa64pfr1(cpu, regs.r64(Regs::k64_AA64PFR0), regs.r64(Regs::k64_AA64PFR1));
    detect_aarch64_features_via_cpuid_aa64isar0_aa64isar1(cpu, regs.r64(Regs::k64_AA64ISAR0), regs.r64(Regs::k64_AA64ISAR1));

    // TODO: AA64ISAR2 should be added when it's provided by NetBSD.
    // detect_aarch64_features_via_cpuid_aa64isar2(cpu, regs.r64(k64_AA64ISAR2)));

    detect_aarch64_features_via_cpuid_aa64mmfr0(cpu, regs.r64(Regs::k64_AA64MMFR0));
    detect_aarch64_features_via_cpuid_aa64mmfr1(cpu, regs.r64(Regs::k64_AA64MMFR1));
    detect_aarch64_features_via_cpuid_aa64mmfr2(cpu, regs.r64(Regs::k64_AA64MMFR2));

    // Only read CPU_ID_AA64ZFR0 when either SVE or SME is available.
    if (cpu.features().arm().has_any(Ext::kSVE, Ext::kSME)) {
      detect_aarch64_features_via_cpuid_aa64zfr0(cpu, regs.r64(Regs::k64_AA64ZFR0));

      // TODO: AA64SMFR0 should be added when it's provided by NetBSD.
      // if (cpu.features().arm().has_sme()) {
      //   detect_aarch64_features_via_cpuid_aa64smfr0(cpu, regs.r64(Regs::k64_kAA64SMFR0));
      // }
    }
  }

  post_process_arm_cpu_info(cpu);
}

// CpuInfo - Detect - ARM - Detect by OpenBSD API That Reads CPUID
// ===============================================================

#elif defined(__OpenBSD__) && ASMJIT_ARCH_ARM >= 64

// Supported CPUID registers on OpenBSD (CTL_MACHDEP definitions):
//   - https://github.com/openbsd/src/blob/master/sys/arch/arm64/include/cpu.h
enum class OpenBSDAArch64CPUID {
  kAA64ISAR0 = 2,
  kAA64ISAR1 = 3,
  kAA64ISAR2 = 4,
  kAA64MMFR0 = 5,
  kAA64MMFR1 = 6,
  kAA64MMFR2 = 7,
  kAA64PFR0 = 8,
  kAA64PFR1 = 9,
  kAA64SMFR0 = 10,
  kAA64ZFR0 = 11
};

static uint64_t openbsd_read_aarch64_cpuid(OpenBSDAArch64CPUID id) noexcept {
  uint64_t bits = 0;
  size_t size = sizeof(bits);
  int name[2] = { CTL_MACHDEP, int(id) };

  return (sysctl(name, 2, &bits, &size, NULL, 0) < 0) ? uint64_t(0) : bits;
}

static ASMJIT_FAVOR_SIZE void detect_arm_cpu(CpuInfo& cpu) noexcept {
  using ID = OpenBSDAArch64CPUID;

  populate_base_arm_features(cpu);

  detect_aarch64_features_via_cpuid_aa64pfr0_aa64pfr1(cpu, openbsd_read_aarch64_cpuid(ID::kAA64PFR0), openbsd_read_aarch64_cpuid(ID::kAA64PFR1));
  detect_aarch64_features_via_cpuid_aa64isar0_aa64isar1(cpu, openbsd_read_aarch64_cpuid(ID::kAA64ISAR0), openbsd_read_aarch64_cpuid(ID::kAA64ISAR1));
  detect_aarch64_features_via_cpuid_aa64isar2(cpu, openbsd_read_aarch64_cpuid(ID::kAA64ISAR2));
  detect_aarch64_features_via_cpuid_aa64mmfr0(cpu, openbsd_read_aarch64_cpuid(ID::kAA64MMFR0));
  detect_aarch64_features_via_cpuid_aa64mmfr1(cpu, openbsd_read_aarch64_cpuid(ID::kAA64MMFR1));
  detect_aarch64_features_via_cpuid_aa64mmfr2(cpu, openbsd_read_aarch64_cpuid(ID::kAA64MMFR2));

  // Only read CPU_ID_AA64ZFR0 when either SVE or SME is available.
  if (cpu.features().arm().has_any(Ext::kSVE, Ext::kSME)) {
    detect_aarch64_features_via_cpuid_aa64zfr0(cpu, openbsd_read_aarch64_cpuid(ID::kAA64ZFR0));

    if (cpu.features().arm().has_sme()) {
      detect_aarch64_features_via_cpuid_aa64smfr0(cpu, openbsd_read_aarch64_cpuid(ID::kAA64SMFR0));
    }
  }

  post_process_arm_cpu_info(cpu);
}

// CpuInfo - Detect - ARM - Detect by Apple API (sysctlbyname)
// ===========================================================

#elif defined(__APPLE__)

enum class AppleFeatureType : uint8_t {
  kHWOptional,
  kHWOptionalArmFEAT
};

struct AppleFeatureMapping {
  AppleFeatureType type;
  char name[18];
  uint8_t feature_id;
};

template<typename T>
static inline bool invoke_sysctl_by_name(const char* sysctl_name, T* dst, size_t size = sizeof(T)) noexcept {
  return sysctlbyname(sysctl_name, dst, &size, nullptr, 0) == 0;
}

static ASMJIT_FAVOR_SIZE long apple_detect_aarch64_feature_via_sysctl(AppleFeatureType type, const char* feature_name) noexcept {
  static const char hw_optional_prefix[] = "hw.optional.";
  static const char hw_optional_arm_feat_prefix[] = "hw.optional.arm.FEAT_";

  char sysctl_name[128];

  const char* prefix = type == AppleFeatureType::kHWOptional ? hw_optional_prefix : hw_optional_arm_feat_prefix;
  size_t prefix_size = (type == AppleFeatureType::kHWOptional ? sizeof(hw_optional_prefix) : sizeof(hw_optional_arm_feat_prefix)) - 1u;
  size_t feature_name_size = strlen(feature_name);

  if (feature_name_size < 128 - prefix_size) {
    memcpy(sysctl_name, prefix, prefix_size);
    memcpy(sysctl_name + prefix_size, feature_name, feature_name_size + 1u); // Include NULL terminator.

    long val = 0;
    if (invoke_sysctl_by_name<long>(sysctl_name, &val)) {
      return val;
    }
  }

  return 0;
}

static ASMJIT_FAVOR_SIZE void apple_detect_aarch64_features_via_sysctl(CpuInfo& cpu) noexcept {
  using FT = AppleFeatureType;

  // Based on:
  //   - https://developer.apple.com/documentation/kernel/1387446-sysctlbyname/determining_instruction_set_characteristics
  static const AppleFeatureMapping mappings[] = {
    // Determine Advanced SIMD and Floating Point Capabilities:
    { FT::kHWOptional       , "AdvSIMD_HPFPCvt", uint8_t(Ext::kFP16CONV)  },
    { FT::kHWOptional       , "neon_hpfp"      , uint8_t(Ext::kFP16CONV)  },
    { FT::kHWOptionalArmFEAT, "BF16"           , uint8_t(Ext::kBF16)      },
    { FT::kHWOptionalArmFEAT, "DotProd"        , uint8_t(Ext::kDOTPROD)   },
    { FT::kHWOptionalArmFEAT, "FCMA"           , uint8_t(Ext::kFCMA)      },
    { FT::kHWOptional       , "armv8_3_compnum", uint8_t(Ext::kFCMA)      },
    { FT::kHWOptionalArmFEAT, "FHM"            , uint8_t(Ext::kFHM)       },
    { FT::kHWOptional       , "armv8_2_fhm"    , uint8_t(Ext::kFHM)       },
    { FT::kHWOptionalArmFEAT, "FP16"           , uint8_t(Ext::kFP16)      },
    { FT::kHWOptional       , "neon_fp16"      , uint8_t(Ext::kFP16)      },
    { FT::kHWOptionalArmFEAT, "FRINTTS"        , uint8_t(Ext::kFRINTTS)   },
    { FT::kHWOptionalArmFEAT, "I8MM"           , uint8_t(Ext::kI8MM)      },
    { FT::kHWOptionalArmFEAT, "JSCVT"          , uint8_t(Ext::kJSCVT)     },
    { FT::kHWOptionalArmFEAT, "RDM"            , uint8_t(Ext::kRDM)       },

    // Determine Integer Capabilities:
    { FT::kHWOptional       , "armv8_crc32"    , uint8_t(Ext::kCRC32)     },
    { FT::kHWOptionalArmFEAT, "FlagM"          , uint8_t(Ext::kFLAGM)     },
    { FT::kHWOptionalArmFEAT, "FlagM2"         , uint8_t(Ext::kFLAGM2)    },

    // Determine Atomic and Memory Ordering Instruction Capabilities:
    { FT::kHWOptionalArmFEAT, "LRCPC"          , uint8_t(Ext::kLRCPC)     },
    { FT::kHWOptionalArmFEAT, "LRCPC2"         , uint8_t(Ext::kLRCPC2)    },
    { FT::kHWOptional       , "armv8_1_atomics", uint8_t(Ext::kLSE)       },
    { FT::kHWOptionalArmFEAT, "LSE"            , uint8_t(Ext::kLSE)       },
    { FT::kHWOptionalArmFEAT, "LSE2"           , uint8_t(Ext::kLSE2)      },

    // Determine Encryption Capabilities:
    { FT::kHWOptionalArmFEAT, "AES"            , uint8_t(Ext::kAES)       },
    { FT::kHWOptionalArmFEAT, "PMULL"          , uint8_t(Ext::kPMULL)     },
    { FT::kHWOptionalArmFEAT, "SHA1"           , uint8_t(Ext::kSHA1)      },
    { FT::kHWOptionalArmFEAT, "SHA256"         , uint8_t(Ext::kSHA256)    },
    { FT::kHWOptionalArmFEAT, "SHA512"         , uint8_t(Ext::kSHA512)    },
    { FT::kHWOptional       , "armv8_2_sha512" , uint8_t(Ext::kSHA512)    },
    { FT::kHWOptionalArmFEAT, "SHA3"           , uint8_t(Ext::kSHA3)      },
    { FT::kHWOptional       , "armv8_2_sha3"   , uint8_t(Ext::kSHA3)      },

    // Determine General Capabilities:
    { FT::kHWOptionalArmFEAT, "BTI"            , uint8_t(Ext::kBTI)       },
    { FT::kHWOptionalArmFEAT, "DPB"            , uint8_t(Ext::kDPB)       },
    { FT::kHWOptionalArmFEAT, "DPB2"           , uint8_t(Ext::kDPB2)      },
    { FT::kHWOptionalArmFEAT, "ECV"            , uint8_t(Ext::kECV)       },
    { FT::kHWOptionalArmFEAT, "SB"             , uint8_t(Ext::kSB)        },
    { FT::kHWOptionalArmFEAT, "SSBS"           , uint8_t(Ext::kSSBS)      }
  };

  for (size_t i = 0; i < ASMJIT_ARRAY_SIZE(mappings); i++) {
    const AppleFeatureMapping& mapping = mappings[i];
    if (!cpu.features().arm().has(mapping.feature_id) && apple_detect_aarch64_feature_via_sysctl(mapping.type, mapping.name)) {
      cpu.features().arm().add(mapping.feature_id);
    }
  }
}

static ASMJIT_FAVOR_SIZE void detect_arm_cpu(CpuInfo& cpu) noexcept {
  cpu._was_detected = true;
  populate_base_arm_features(cpu);

  invoke_sysctl_by_name<uint32_t>("hw.cpufamily", &cpu._family_id);
  invoke_sysctl_by_name<uint32_t>("hw.cachelinesize", &cpu._cache_line_size);
  invoke_sysctl_by_name<uint32_t>("machdep.cpu.logical_per_package", &cpu._max_logical_processors);
  invoke_sysctl_by_name<char>("machdep.cpu.brand_string", cpu._brand.str, sizeof(cpu._brand.str));

  memcpy(cpu._vendor.str, "APPLE", 6);

  bool cpu_features_populated = detect_aarch64_features_via_apple_family_id(cpu);
  if (!cpu_features_populated) {
    apple_detect_aarch64_features_via_sysctl(cpu);
  }
  post_process_arm_cpu_info(cpu);
}

// CpuInfo - Detect - ARM - Detect by Fallback (Using Compiler Flags)
// ==================================================================

#else

#if ASMJIT_ARCH_ARM == 32
  #pragma message("[asmjit] Disabling runtime CPU detection - unsupported OS/CPU combination (Unknown OS with AArch32 CPU)")
#else
  #pragma message("[asmjit] Disabling runtime CPU detection - unsupported OS/CPU combination (Unknown OS with AArch64 CPU)")
#endif

static ASMJIT_FAVOR_SIZE void detect_arm_cpu(CpuInfo& cpu) noexcept {
  populate_base_arm_features(cpu);
  detect_arm_features_via_compiler_flags(cpu);
  post_process_arm_cpu_info(cpu);
}
#endif

static ASMJIT_FAVOR_SIZE CpuHints recalculate_hints(const CpuInfo& cpu_info, const CpuFeatures::ARM& features) noexcept {
  Support::maybe_unused(cpu_info, features);

  // Assume ARM CPUs have fast 32-bit SIMD integer multiplication.
  CpuHints hints = CpuHints::kVecFastIntMul32;

  return hints;
}

} // {arm}

#endif

// CpuInfo - Detect - Host
// =======================

const CpuInfo& CpuInfo::host() noexcept {
  static std::atomic<uint32_t> cpu_info_initialized_flag;
  static CpuInfo cpu_info_global(Globals::NoInit);

  // This should never cause a problem as the resulting information should always
  // be the same. In the worst case it would just be overwritten non-atomically.
  if (!cpu_info_initialized_flag.load(std::memory_order_relaxed)) {
    CpuInfo cpu_info_local;

    cpu_info_local._arch = Arch::kHost;
    cpu_info_local._sub_arch = SubArch::kHost;

#if ASMJIT_ARCH_X86
    x86::detect_x86_cpu(cpu_info_local);
#elif ASMJIT_ARCH_ARM
    arm::detect_arm_cpu(cpu_info_local);
#endif
    cpu_info_local._hw_thread_count = detect_hw_thread_count();
    cpu_info_local.update_hints();

    cpu_info_global = cpu_info_local;
    cpu_info_initialized_flag.store(1, std::memory_order_seq_cst);
  }

  return cpu_info_global;
}

CpuHints CpuInfo::recalculate_hints(const CpuInfo& info, const CpuFeatures& features) noexcept {
#if ASMJIT_ARCH_X86
  return x86::recalculate_hints(info, features.x86());
#elif ASMJIT_ARCH_ARM
  return arm::recalculate_hints(info, features.arm());
#else
  Support::maybe_unused(info, features);
  return CpuHints::kNone;
#endif
}

ASMJIT_END_NAMESPACE