Merge pull request #10876 from shuffle2/cpudetect

CPUDetect: improve win/arm64 support

8 files changed, 397 insertions, 228 deletions

diff --git a/Source/Core/Common/ArmCPUDetect.cpp b/Source/Core/Common/ArmCPUDetect.cpp
index a12deac71a..7fe2d14220 100644
--- a/Source/Core/Common/ArmCPUDetect.cpp
+++ b/Source/Core/Common/ArmCPUDetect.cpp
@@ -9,45 +9,209 @@
 #include <string>
 #include <thread>
 
-#if !defined(_WIN32) && !defined(__APPLE__)
+#ifdef __APPLE__
+#include <sys/sysctl.h>
+#elif defined(_WIN32)
+#include <Windows.h>
+#include <arm64intr.h>
+#else
 #ifndef __FreeBSD__
 #include <asm/hwcap.h>
 #endif
 #include <sys/auxv.h>
-#include <unistd.h>
 #endif
 
 #include <fmt/format.h>
 
 #include "Common/CommonTypes.h"
 #include "Common/FileUtil.h"
+#include "Common/StringUtil.h"
+
+#if defined(__APPLE__) || defined(__FreeBSD__)
+
+static bool SysctlByName(std::string* value, const std::string& name)
+{
+  size_t value_len = 0;
+  if (sysctlbyname(name.c_str(), nullptr, &value_len, nullptr, 0))
+    return false;
 
-#ifndef WIN32
+  value->resize(value_len);
+  if (sysctlbyname(name.c_str(), value->data(), &value_len, nullptr, 0))
+    return false;
 
-const char procfile[] = "/proc/cpuinfo";
+  TruncateToCString(value);
+  return true;
+}
+
+#endif
 
-static std::string GetCPUString()
+#if defined(_WIN32)
+
+static constexpr char SUBKEY_CORE0[] = R"(HARDWARE\DESCRIPTION\System\CentralProcessor\0)";
+
+// Identifier: human-readable version of CPUID
+// ProcessorNameString: marketing name of the processor
+// VendorIdentifier: vendor company name
+// There are some other maybe-interesting values nearby, BIOS info etc.
+static bool ReadProcessorString(std::string* value, const std::string& name)
 {
-  const std::string marker = "Hardware\t: ";
-  std::string cpu_string = "Unknown";
+  const DWORD flags = RRF_RT_REG_SZ | RRF_NOEXPAND;
+  DWORD value_len = 0;
+  auto status = RegGetValueA(HKEY_LOCAL_MACHINE, SUBKEY_CORE0, name.c_str(), flags, nullptr,
+                             nullptr, &value_len);
+  if (status != ERROR_SUCCESS && status != ERROR_MORE_DATA)
+    return false;
 
-  std::string line;
+  value->resize(value_len);
+  status = RegGetValueA(HKEY_LOCAL_MACHINE, SUBKEY_CORE0, name.c_str(), flags, nullptr,
+                        value->data(), &value_len);
+  if (status != ERROR_SUCCESS)
+  {
+    value->clear();
+    return false;
+  }
+
+  TruncateToCString(value);
+  return true;
+}
+
+// Read cached register values from the registry
+static bool ReadPrivilegedCPReg(u64* value, u32 reg)
+{
+  DWORD value_len = sizeof(*value);
+  // Not sure if the value name is padded or not
+  return RegGetValueA(HKEY_LOCAL_MACHINE, SUBKEY_CORE0, fmt::format("CP {:x}", reg).c_str(),
+                      RRF_RT_REG_QWORD, nullptr, value, &value_len) == ERROR_SUCCESS;
+}
+
+static bool Read_MIDR_EL1(u64* value)
+{
+  return ReadPrivilegedCPReg(value, ARM64_SYSREG(0b11, 0, 0, 0b0000, 0));
+}
+
+static bool Read_ID_AA64ISAR0_EL1(u64* value)
+{
+  return ReadPrivilegedCPReg(value, ARM64_SYSREG(0b11, 0, 0, 0b0110, 0));
+}
+
+static bool Read_ID_AA64MMFR1_EL1(u64* value)
+{
+  return ReadPrivilegedCPReg(value, ARM64_SYSREG(0b11, 0, 0, 0b0111, 1));
+}
+
+#endif
+
+#if defined(__linux__)
+
+static bool ReadDeviceTree(std::string* value, const std::string& name)
+{
+  const std::string path = std::string("/proc/device-tree/") + name;
   std::ifstream file;
-  File::OpenFStream(file, procfile, std::ios_base::in);
+  File::OpenFStream(file, path.c_str(), std::ios_base::in);
+  if (!file)
+    return false;
 
+  file >> *value;
+  return true;
+}
+
+static std::string ReadCpuinfoField(const std::string& field)
+{
+  std::string line;
+  std::ifstream file;
+  File::OpenFStream(file, "/proc/cpuinfo", std::ios_base::in);
   if (!file)
-    return cpu_string;
+    return {};
 
   while (std::getline(file, line))
   {
-    if (line.find(marker) != std::string::npos)
-    {
-      cpu_string = line.substr(marker.length());
-      break;
-    }
+    if (!StringBeginsWith(line, field))
+      continue;
+    auto non_tab = line.find_first_not_of("\t", field.length());
+    if (non_tab == line.npos)
+      continue;
+    if (line[non_tab] != ':')
+      continue;
+    auto value_start = line.find_first_not_of(" ", non_tab + 1);
+    if (value_start == line.npos)
+      continue;
+    return line.substr(value_start);
   }
+  return {};
+}
 
-  return cpu_string;
+static bool Read_MIDR_EL1_Sysfs(u64* value)
+{
+  std::ifstream file;
+  File::OpenFStream(file, "/sys/devices/system/cpu/cpu0/regs/identification/midr_el1",
+                    std::ios_base::in);
+  if (!file)
+    return false;
+
+  file >> std::hex >> *value;
+  return true;
+}
+
+#endif
+
+#if defined(__linux__) || defined(__FreeBSD__)
+
+static u32 ReadHwCap(u32 type)
+{
+#if defined(__linux__)
+  return getauxval(type);
+#elif defined(__FreeBSD__)
+  u_long hwcap = 0;
+  elf_aux_info(type, &hwcap, sizeof(hwcap));
+  return hwcap;
+#endif
+}
+
+// For "Direct" reads, value gets filled via emulation, hence:
+// "there is no guarantee that the value reflects the processor that it is currently executing on"
+// On big.LITTLE systems, the value may be unrelated to the core this is invoked on, and unless
+// other measures are taken, executing the instruction may cause the caller to be switched onto a
+// different core when it resumes (and of course, caller could be preempted at any other time as
+// well).
+static inline u64 Read_MIDR_EL1_Direct()
+{
+  u64 value;
+  __asm__ __volatile__("mrs %0, MIDR_EL1" : "=r"(value));
+  return value;
+}
+
+static bool Read_MIDR_EL1(u64* value)
+{
+#ifdef __linux__
+  if (Read_MIDR_EL1_Sysfs(value))
+    return true;
+#endif
+
+  bool id_reg_user_access = ReadHwCap(AT_HWCAP) & HWCAP_CPUID;
+#ifdef __FreeBSD__
+  // FreeBSD kernel has support but doesn't seem to indicate it?
+  // see user_mrs_handler
+  id_reg_user_access = true;
+#endif
+  if (!id_reg_user_access)
+    return false;
+  *value = Read_MIDR_EL1_Direct();
+  return true;
+}
+
+#endif
+
+#ifndef __APPLE__
+
+static std::string MIDRToString(u64 midr)
+{
+  u8 implementer = (midr >> 24) & 0xff;
+  u8 variant = (midr >> 20) & 0xf;
+  u8 arch = (midr >> 16) & 0xf;
+  u16 part_num = (midr >> 4) & 0xfff;
+  u8 revision = midr & 0xf;
+  return fmt::format("{:02X}:{:X}:{:04b}:{:03X}:{:X}", implementer, variant, arch, part_num,
+                     revision);
 }
 
 #endif
@@ -59,82 +223,100 @@ CPUInfo::CPUInfo()
   Detect();
 }
 
-// Detects the various CPU features
 void CPUInfo::Detect()
 {
-  // Set some defaults here
-  HTT = false;
-  OS64bit = true;
-  CPU64bit = true;
-  Mode64bit = true;
   vendor = CPUVendor::ARM;
   bFMA = true;
   bFlushToZero = true;
-  bAFP = false;
+
+  num_cores = std::max(static_cast<int>(std::thread::hardware_concurrency()), 1);
 
 #ifdef __APPLE__
-  num_cores = std::thread::hardware_concurrency();
+  SysctlByName(&model_name, "machdep.cpu.brand_string");
 
   // M-series CPUs have all of these
-  bFP = true;
-  bASIMD = true;
+  // Apparently the world has accepted that these can be assumed supported "for all time".
+  // see https://github.com/golang/go/issues/42747
   bAES = true;
   bSHA1 = true;
   bSHA2 = true;
   bCRC32 = true;
 #elif defined(_WIN32)
-  num_cores = std::thread::hardware_concurrency();
-
-  // Windows does not provide any mechanism for querying the system registers on ARMv8, unlike Linux
-  // which traps the register reads and emulates them in the kernel. There are environment variables
-  // containing some of the CPU-specific values, which we could use for a lookup table in the
-  // future. For now, assume all features are present as all known devices which are Windows-on-ARM
-  // compatible also support these extensions.
-  bFP = true;
-  bASIMD = true;
-  bAES = true;
-  bCRC32 = true;
-  bSHA1 = true;
-  bSHA2 = true;
-#else
-  // Get the information about the CPU
-  num_cores = sysconf(_SC_NPROCESSORS_CONF);
-  strncpy(cpu_string, GetCPUString().c_str(), sizeof(cpu_string));
+  // NOTE All this info is from cpu core 0 only.
 
-#ifdef __FreeBSD__
-  u_long hwcaps = 0;
-  elf_aux_info(AT_HWCAP, &hwcaps, sizeof(u_long));
+  ReadProcessorString(&model_name, "ProcessorNameString");
+
+  u64 reg = 0;
+  // Attempt to be forward-compatible: perform inverted check against disabled feature states.
+  if (Read_ID_AA64ISAR0_EL1(&reg))
+  {
+    bAES = ((reg >> 4) & 0xf) != 0;
+    bSHA1 = ((reg >> 8) & 0xf) != 0;
+    bSHA2 = ((reg >> 12) & 0xf) != 0;
+    bCRC32 = ((reg >> 16) & 0xf) != 0;
+  }
+  if (Read_ID_AA64MMFR1_EL1(&reg))
+  {
+    // Introduced in Armv8.7, where AFP must be supported if AdvSIMD and FP both are.
+    bAFP = ((reg >> 44) & 0xf) != 0;
+  }
+  // Pre-decoded MIDR_EL1 could be read with ReadProcessorString(.., "Identifier"),
+  // but we want format to match across all platforms where possible.
+  if (Read_MIDR_EL1(&reg))
+  {
+    cpu_id = MIDRToString(reg);
+  }
 #else
-  unsigned long hwcaps = getauxval(AT_HWCAP);
+  // Linux, Android, and FreeBSD
+
+#if defined(__FreeBSD__)
+  SysctlByName(&model_name, "hw.model");
+#elif defined(__linux__)
+  if (!ReadDeviceTree(&model_name, "model"))
+  {
+    // This doesn't seem to work on modern arm64 kernels
+    model_name = ReadCpuinfoField("Hardware");
+  }
 #endif
-  bFP = hwcaps & HWCAP_FP;
-  bASIMD = hwcaps & HWCAP_ASIMD;
-  bAES = hwcaps & HWCAP_AES;
-  bCRC32 = hwcaps & HWCAP_CRC32;
-  bSHA1 = hwcaps & HWCAP_SHA1;
-  bSHA2 = hwcaps & HWCAP_SHA2;
+
+  const u32 hwcap = ReadHwCap(AT_HWCAP);
+  bAES = hwcap & HWCAP_AES;
+  bCRC32 = hwcap & HWCAP_CRC32;
+  bSHA1 = hwcap & HWCAP_SHA1;
+  bSHA2 = hwcap & HWCAP_SHA2;
+
+#if defined(AT_HWCAP2) && defined(HWCAP2_AFP)
+  const u32 hwcap2 = ReadHwCap(AT_HWCAP2);
+  bAFP = hwcap2 & HWCAP2_AFP;
 #endif
+
+  u64 midr = 0;
+  if (Read_MIDR_EL1(&midr))
+  {
+    cpu_id = MIDRToString(midr);
+  }
+#endif
+
+  model_name = ReplaceAll(model_name, ",", "_");
+  cpu_id = ReplaceAll(cpu_id, ",", "_");
 }
 
-// Turn the CPU info into a string we can show
 std::string CPUInfo::Summarize()
 {
-  std::string sum;
-  if (num_cores == 1)
-    sum = fmt::format("{}, 1 core", cpu_string);
-  else
-    sum = fmt::format("{}, {} cores", cpu_string, num_cores);
+  std::vector<std::string> sum;
+  sum.push_back(model_name);
+  sum.push_back(cpu_id);
 
+  if (bAFP)
+    sum.push_back("AFP");
   if (bAES)
-    sum += ", AES";
+    sum.push_back("AES");
   if (bCRC32)
-    sum += ", CRC32";
+    sum.push_back("CRC32");
   if (bSHA1)
-    sum += ", SHA1";
+    sum.push_back("SHA1");
   if (bSHA2)
-    sum += ", SHA2";
-  if (CPU64bit)
-    sum += ", 64-bit";
+    sum.push_back("SHA2");
 
-  return sum;
+  return JoinStrings(sum, ",");
 }
diff --git a/Source/Core/Common/CPUDetect.h b/Source/Core/Common/CPUDetect.h
index db377a2006..5a4fac2364 100644
--- a/Source/Core/Common/CPUDetect.h
+++ b/Source/Core/Common/CPUDetect.h
@@ -16,62 +16,52 @@ enum class CPUVendor
 
 struct CPUInfo
 {
-  CPUVendor vendor = CPUVendor::Intel;
+  CPUVendor vendor = CPUVendor::Other;
 
-  char cpu_string[0x41] = {};
-  char brand_string[0x21] = {};
-  bool OS64bit = false;
-  bool CPU64bit = false;
-  bool Mode64bit = false;
+  std::string cpu_id;
+  std::string model_name;
 
   bool HTT = false;
   int num_cores = 0;
 
-  bool bSSE = false;
-  bool bSSE2 = false;
   bool bSSE3 = false;
   bool bSSSE3 = false;
-  bool bPOPCNT = false;
   bool bSSE4_1 = false;
   bool bSSE4_2 = false;
   bool bLZCNT = false;
-  bool bSSE4A = false;
   bool bAVX = false;
-  bool bAVX2 = false;
   bool bBMI1 = false;
   bool bBMI2 = false;
-  // PDEP and PEXT are ridiculously slow on AMD Zen1, Zen1+ and Zen2 (Family 23)
-  bool bFastBMI2 = false;
+  // PDEP and PEXT are ridiculously slow on AMD Zen1, Zen1+ and Zen2 (Family 17h)
+  bool bBMI2FastParallelBitOps = false;
   bool bFMA = false;
   bool bFMA4 = false;
   bool bAES = false;
-  // FXSAVE/FXRSTOR
-  bool bFXSR = false;
   bool bMOVBE = false;
   // This flag indicates that the hardware supports some mode
   // in which denormal inputs _and_ outputs are automatically set to (signed) zero.
   bool bFlushToZero = false;
-  bool bLAHFSAHF64 = false;
-  bool bLongMode = false;
   bool bAtom = false;
-  bool bZen1p2 = false;
-
-  // ARMv8 specific
-  bool bFP = false;
-  bool bASIMD = false;
   bool bCRC32 = false;
   bool bSHA1 = false;
   bool bSHA2 = false;
+
+  // ARMv8 specific
   bool bAFP = false;  // Alternate floating-point behavior
 
   // Call Detect()
   explicit CPUInfo();
 
-  // Turn the CPU info into a string we can show
+  // The returned string consists of "<model_name>,<cpu_id>,<flag...>"
+  // Where:
+  //  model_name and cpud_id may be zero-length
+  //  model_name is human-readable marketing name
+  //  cpu_id is ':'-delimited string of id info
+  //  flags are optionally included if the related feature is supported and reporting its enablement
+  //  seems useful to report
   std::string Summarize();
 
 private:
-  // Detects the various CPU features
   void Detect();
 };
 
diff --git a/Source/Core/Common/StringUtil.cpp b/Source/Core/Common/StringUtil.cpp
index 4a5f21ddc2..849e83815c 100644
--- a/Source/Core/Common/StringUtil.cpp
+++ b/Source/Core/Common/StringUtil.cpp
@@ -259,6 +259,13 @@ void ReplaceBreaksWithSpaces(std::string& str)
   std::replace(str.begin(), str.end(), '\n', ' ');
 }
 
+void TruncateToCString(std::string* s)
+{
+  const size_t terminator = s->find_first_of('\0');
+  if (terminator != s->npos)
+    s->resize(terminator);
+}
+
 bool TryParse(const std::string& str, bool* const output)
 {
   float value;
diff --git a/Source/Core/Common/StringUtil.h b/Source/Core/Common/StringUtil.h
index 0df7b8d119..7cd8b691bc 100644
--- a/Source/Core/Common/StringUtil.h
+++ b/Source/Core/Common/StringUtil.h
@@ -54,6 +54,8 @@ std::string ReplaceAll(std::string result, std::string_view src, std::string_vie
 
 void ReplaceBreaksWithSpaces(std::string& str);
 
+void TruncateToCString(std::string* s);
+
 bool TryParse(const std::string& str, bool* output);
 
 template <typename T, std::enable_if_t<std::is_integral_v<T> || std::is_enum_v<T>>* = nullptr>
diff --git a/Source/Core/Common/x64CPUDetect.cpp b/Source/Core/Common/x64CPUDetect.cpp
index 5f90246302..dedf02c05e 100644
--- a/Source/Core/Common/x64CPUDetect.cpp
+++ b/Source/Core/Common/x64CPUDetect.cpp
@@ -11,9 +11,12 @@
 #include <string>
 #include <thread>
 
+#include <fmt/format.h>
+
 #include "Common/CommonTypes.h"
 #include "Common/Intrinsics.h"
 #include "Common/MsgHandler.h"
+#include "Common/StringUtil.h"
 
 #ifndef _WIN32
 
@@ -38,20 +41,23 @@ static inline void __cpuidex(int info[4], int function_id, int subfunction_id)
 #endif
 }
 
-static inline void __cpuid(int info[4], int function_id)
+constexpr u32 XCR_XFEATURE_ENABLED_MASK = 0;
+
+static u64 xgetbv(u32 index)
 {
-  return __cpuidex(info, function_id, 0);
+  u32 eax, edx;
+  __asm__ __volatile__("xgetbv" : "=a"(eax), "=d"(edx) : "c"(index));
+  return ((u64)edx << 32) | eax;
 }
 
-#endif  // ifndef _WIN32
+#else
 
-#ifdef _WIN32
+constexpr u32 XCR_XFEATURE_ENABLED_MASK = _XCR_XFEATURE_ENABLED_MASK;
 
 static u64 xgetbv(u32 index)
 {
   return _xgetbv(index);
 }
-constexpr u32 XCR_XFEATURE_ENABLED_MASK = _XCR_XFEATURE_ENABLED_MASK;
 
 static void WarnIfRunningUnderEmulation()
 {
@@ -76,16 +82,20 @@ static void WarnIfRunningUnderEmulation()
                  "Please run the ARM64 build of Dolphin for a better experience.");
 }
 
-#else
+#endif  // ifdef _WIN32
 
-static u64 xgetbv(u32 index)
+struct CPUIDResult
 {
-  u32 eax, edx;
-  __asm__ __volatile__("xgetbv" : "=a"(eax), "=d"(edx) : "c"(index));
-  return ((u64)edx << 32) | eax;
+  u32 eax{}, ebx{}, ecx{}, edx{};
+};
+static_assert(sizeof(CPUIDResult) == sizeof(u32) * 4);
+
+static inline CPUIDResult cpuid(int function_id, int subfunction_id = 0)
+{
+  CPUIDResult info;
+  __cpuidex((int*)&info, function_id, subfunction_id);
+  return info;
 }
-constexpr u32 XCR_XFEATURE_ENABLED_MASK = 0;
-#endif  // ifdef _WIN32
 
 CPUInfo cpu_info;
 
@@ -94,196 +104,174 @@ CPUInfo::CPUInfo()
   Detect();
 }
 
-// Detects the various CPU features
 void CPUInfo::Detect()
 {
 #ifdef _WIN32
   WarnIfRunningUnderEmulation();
 #endif
 
-#ifdef _M_X86_64
-  Mode64bit = true;
-  OS64bit = true;
-#endif
-  num_cores = 1;
-
-  // Set obvious defaults, for extra safety
-  if (Mode64bit)
-  {
-    bSSE = true;
-    bSSE2 = true;
-    bLongMode = true;
-  }
+  // This should be much more reliable and easier than trying to get the number of cores out of the
+  // CPUID data ourselves.
+  num_cores = std::max(static_cast<int>(std::thread::hardware_concurrency()), 1);
 
   // Assume CPU supports the CPUID instruction. Those that don't can barely
-  // boot modern OS:es anyway.
-  int cpu_id[4];
-
-  // Detect CPU's CPUID capabilities, and grab CPU string
-  __cpuid(cpu_id, 0x00000000);
-  u32 max_std_fn = cpu_id[0];  // EAX
-  std::memcpy(&brand_string[0], &cpu_id[1], sizeof(int));
-  std::memcpy(&brand_string[4], &cpu_id[3], sizeof(int));
-  std::memcpy(&brand_string[8], &cpu_id[2], sizeof(int));
-  __cpuid(cpu_id, 0x80000000);
-  u32 max_ex_fn = cpu_id[0];
-  if (!strcmp(brand_string, "GenuineIntel"))
+  // boot modern OS anyway.
+
+  // Detect CPU's CPUID capabilities and grab vendor string.
+  auto info = cpuid(0);
+  const u32 func_id_max = info.eax;
+
+  std::string vendor_id;
+  vendor_id.resize(sizeof(u32) * 3);
+  std::memcpy(&vendor_id[0], &info.ebx, sizeof(u32));
+  std::memcpy(&vendor_id[4], &info.edx, sizeof(u32));
+  std::memcpy(&vendor_id[8], &info.ecx, sizeof(u32));
+  TruncateToCString(&vendor_id);
+  if (vendor_id == "GenuineIntel")
     vendor = CPUVendor::Intel;
-  else if (!strcmp(brand_string, "AuthenticAMD"))
+  else if (vendor_id == "AuthenticAMD")
     vendor = CPUVendor::AMD;
   else
     vendor = CPUVendor::Other;
 
-  // Set reasonable default brand string even if brand string not available.
-  strcpy(cpu_string, brand_string);
-
   // Detect family and other misc stuff.
-  bool ht = false;
-  HTT = ht;
-  if (max_std_fn >= 1)
+  bool is_amd_family_17 = false;
+  bool has_sse = false;
+  if (func_id_max >= 1)
   {
-    __cpuid(cpu_id, 0x00000001);
-    int family = ((cpu_id[0] >> 8) & 0xf) + ((cpu_id[0] >> 20) & 0xff);
-    int model = ((cpu_id[0] >> 4) & 0xf) + ((cpu_id[0] >> 12) & 0xf0);
+    info = cpuid(1);
+    const u32 version = info.eax;
+    const u32 family = ((version >> 8) & 0xf) + ((version >> 20) & 0xff);
+    const u32 model = ((version >> 4) & 0xf) + ((version >> 12) & 0xf0);
+    const u32 stepping = version & 0xf;
+
+    cpu_id = fmt::format("{:02X}:{:02X}:{:X}", family, model, stepping);
+
     // Detect people unfortunate enough to be running Dolphin on an Atom
-    if (family == 6 &&
+    if (vendor == CPUVendor::Intel && family == 6 &&
         (model == 0x1C || model == 0x26 || model == 0x27 || model == 0x35 || model == 0x36 ||
          model == 0x37 || model == 0x4A || model == 0x4D || model == 0x5A || model == 0x5D))
       bAtom = true;
+
     // Detect AMD Zen1, Zen1+ and Zen2
-    if (family == 23)
-      bZen1p2 = true;
-    ht = (cpu_id[3] >> 28) & 1;
+    if (vendor == CPUVendor::AMD && family == 0x17)
+      is_amd_family_17 = true;
 
     // AMD CPUs before Zen faked this flag and didn't actually
     // implement simultaneous multithreading (SMT; Intel calls it HTT)
     // but rather some weird middle-ground between 1-2 cores
-    HTT = ht && (vendor == CPUVendor::Intel || family >= 23);
+    const bool ht = (info.edx >> 28) & 1;
+    HTT = ht && (vendor == CPUVendor::Intel || (vendor == CPUVendor::AMD && family >= 0x17));
 
-    if ((cpu_id[3] >> 25) & 1)
-      bSSE = true;
-    if ((cpu_id[3] >> 26) & 1)
-      bSSE2 = true;
-    if ((cpu_id[2]) & 1)
+    if ((info.edx >> 25) & 1)
+      has_sse = true;
+    if (info.ecx & 1)
       bSSE3 = true;
-    if ((cpu_id[2] >> 9) & 1)
+    if ((info.ecx >> 9) & 1)
       bSSSE3 = true;
-    if ((cpu_id[2] >> 19) & 1)
+    if ((info.ecx >> 19) & 1)
       bSSE4_1 = true;
-    if ((cpu_id[2] >> 20) & 1)
+    if ((info.ecx >> 20) & 1)
       bSSE4_2 = true;
-    if ((cpu_id[2] >> 22) & 1)
+    if ((info.ecx >> 22) & 1)
       bMOVBE = true;
-    if ((cpu_id[2] >> 25) & 1)
+    if ((info.ecx >> 25) & 1)
       bAES = true;
 
-    if ((cpu_id[3] >> 24) & 1)
-    {
-      // We can use FXSAVE.
-      bFXSR = true;
-    }
-
     // AVX support requires 3 separate checks:
     //  - Is the AVX bit set in CPUID?
     //  - Is the XSAVE bit set in CPUID?
     //  - XGETBV result has the XCR bit set.
-    if (((cpu_id[2] >> 28) & 1) && ((cpu_id[2] >> 27) & 1))
+    if (((info.ecx >> 28) & 1) && ((info.ecx >> 27) & 1))
     {
-      if ((xgetbv(XCR_XFEATURE_ENABLED_MASK) & 0x6) == 0x6)
+      // Check that XSAVE can be used for SSE and AVX
+      if ((xgetbv(XCR_XFEATURE_ENABLED_MASK) & 0b110) == 0b110)
       {
         bAVX = true;
-        if ((cpu_id[2] >> 12) & 1)
+        if ((info.ecx >> 12) & 1)
           bFMA = true;
       }
     }
 
-    if (max_std_fn >= 7)
+    if (func_id_max >= 7)
     {
-      __cpuidex(cpu_id, 0x00000007, 0x00000000);
-      // careful; we can't enable AVX2 unless the XSAVE/XGETBV checks above passed
-      if ((cpu_id[1] >> 5) & 1)
-        bAVX2 = bAVX;
-      if ((cpu_id[1] >> 3) & 1)
+      info = cpuid(7);
+      if ((info.ebx >> 3) & 1)
         bBMI1 = true;
-      if ((cpu_id[1] >> 8) & 1)
+      if ((info.ebx >> 8) & 1)
         bBMI2 = true;
+      if ((info.ebx >> 29) & 1)
+        bSHA1 = bSHA2 = true;
     }
   }
 
-  bFlushToZero = bSSE;
-  bFastBMI2 = bBMI2 && !bZen1p2;
-
-  if (max_ex_fn >= 0x80000004)
+  info = cpuid(0x80000000);
+  const u32 ext_func_id_max = info.eax;
+  if (ext_func_id_max >= 0x80000004)
   {
     // Extract CPU model string
-    __cpuid(cpu_id, 0x80000002);
-    memcpy(cpu_string, cpu_id, sizeof(cpu_id));
-    __cpuid(cpu_id, 0x80000003);
-    memcpy(cpu_string + 16, cpu_id, sizeof(cpu_id));
-    __cpuid(cpu_id, 0x80000004);
-    memcpy(cpu_string + 32, cpu_id, sizeof(cpu_id));
+    model_name.resize(sizeof(info) * 3);
+    for (u32 i = 0; i < 3; i++)
+    {
+      info = cpuid(0x80000002 + i);
+      memcpy(&model_name[sizeof(info) * i], &info, sizeof(info));
+    }
+    TruncateToCString(&model_name);
+    model_name = StripSpaces(model_name);
   }
-  if (max_ex_fn >= 0x80000001)
+  if (ext_func_id_max >= 0x80000001)
   {
     // Check for more features.
-    __cpuid(cpu_id, 0x80000001);
-    if (cpu_id[2] & 1)
-      bLAHFSAHF64 = true;
-    if ((cpu_id[2] >> 5) & 1)
+    info = cpuid(0x80000001);
+    if ((info.ecx >> 5) & 1)
       bLZCNT = true;
-    if ((cpu_id[2] >> 16) & 1)
+    if ((info.ecx >> 16) & 1)
       bFMA4 = true;
-    if ((cpu_id[3] >> 29) & 1)
-      bLongMode = true;
   }
 
-  // this should be much more reliable and easier
-  // than trying to get the number of cores out of the CPUID data
-  // ourselves
-  num_cores = std::max(std::thread::hardware_concurrency(), 1u);
+  // Computed flags
+  bFlushToZero = has_sse;
+  bBMI2FastParallelBitOps = bBMI2 && !is_amd_family_17;
+  bCRC32 = bSSE4_2;
+
+  model_name = ReplaceAll(model_name, ",", "_");
+  cpu_id = ReplaceAll(cpu_id, ",", "_");
 }
 
-// Turn the CPU info into a string we can show
 std::string CPUInfo::Summarize()
 {
-  std::string sum(cpu_string);
-  sum += " (";
-  sum += brand_string;
-  sum += ")";
-
-  if (bSSE)
-    sum += ", SSE";
-  if (bSSE2)
-  {
-    sum += ", SSE2";
-    if (!bFlushToZero)
-      sum += " (but not DAZ!)";
-  }
+  std::vector<std::string> sum;
+  sum.push_back(model_name);
+  sum.push_back(cpu_id);
+
   if (bSSE3)
-    sum += ", SSE3";
+    sum.push_back("SSE3");
   if (bSSSE3)
-    sum += ", SSSE3";
+    sum.push_back("SSSE3");
   if (bSSE4_1)
-    sum += ", SSE4.1";
+    sum.push_back("SSE4.1");
   if (bSSE4_2)
-    sum += ", SSE4.2";
+    sum.push_back("SSE4.2");
   if (HTT)
-    sum += ", HTT";
+    sum.push_back("HTT");
   if (bAVX)
-    sum += ", AVX";
-  if (bAVX2)
-    sum += ", AVX2";
+    sum.push_back("AVX");
   if (bBMI1)
-    sum += ", BMI1";
+    sum.push_back("BMI1");
   if (bBMI2)
-    sum += ", BMI2";
+    sum.push_back("BMI2");
   if (bFMA)
-    sum += ", FMA";
-  if (bAES)
-    sum += ", AES";
+    sum.push_back("FMA");
   if (bMOVBE)
-    sum += ", MOVBE";
-  if (bLongMode)
-    sum += ", 64-bit support";
-  return sum;
+    sum.push_back("MOVBE");
+  if (bAES)
+    sum.push_back("AES");
+  if (bCRC32)
+    sum.push_back("CRC32");
+  if (bSHA1)
+    sum.push_back("SHA1");
+  if (bSHA2)
+    sum.push_back("SHA2");
+
+  return JoinStrings(sum, ",");
 }
diff --git a/Source/Core/Core/PowerPC/Jit64Common/Jit64AsmCommon.cpp b/Source/Core/Core/PowerPC/Jit64Common/Jit64AsmCommon.cpp
index 124218fb7f..1269b0ec23 100644
--- a/Source/Core/Core/PowerPC/Jit64Common/Jit64AsmCommon.cpp
+++ b/Source/Core/Core/PowerPC/Jit64Common/Jit64AsmCommon.cpp
@@ -65,7 +65,7 @@ void CommonAsmRoutines::GenConvertDoubleToSingle()
 
   // Don't Denormalize
 
-  if (cpu_info.bFastBMI2)
+  if (cpu_info.bBMI2FastParallelBitOps)
   {
     // Extract bits 0-1 and 5-34
     MOV(64, R(RSCRATCH), Imm64(0xc7ffffffe0000000));
diff --git a/Source/Core/VideoCommon/VertexLoaderX64.cpp b/Source/Core/VideoCommon/VertexLoaderX64.cpp
index 63752429aa..27e24e6f28 100644
--- a/Source/Core/VideoCommon/VertexLoaderX64.cpp
+++ b/Source/Core/VideoCommon/VertexLoaderX64.cpp
@@ -299,7 +299,7 @@ void VertexLoaderX64::ReadColor(OpArg data, VertexComponentFormat attribute, Col
     //                   RRRRRGGG GGGBBBBB
     // AAAAAAAA BBBBBBBB GGGGGGGG RRRRRRRR
     LoadAndSwap(16, scratch1, data);
-    if (cpu_info.bBMI1 && cpu_info.bFastBMI2)
+    if (cpu_info.bBMI1 && cpu_info.bBMI2FastParallelBitOps)
     {
       MOV(32, R(scratch2), Imm32(0x07C3F7C0));
       PDEP(32, scratch3, scratch1, R(scratch2));
@@ -339,7 +339,7 @@ void VertexLoaderX64::ReadColor(OpArg data, VertexComponentFormat attribute, Col
     //                   RRRRGGGG BBBBAAAA
     // AAAAAAAA BBBBBBBB GGGGGGGG RRRRRRRR
     LoadAndSwap(16, scratch1, data);
-    if (cpu_info.bFastBMI2)
+    if (cpu_info.bBMI2FastParallelBitOps)
     {
       MOV(32, R(scratch2), Imm32(0x0F0F0F0F));
       PDEP(32, scratch1, scratch1, R(scratch2));
@@ -368,7 +368,7 @@ void VertexLoaderX64::ReadColor(OpArg data, VertexComponentFormat attribute, Col
     // AAAAAAAA BBBBBBBB GGGGGGGG RRRRRRRR
     data.AddMemOffset(-1);  // subtract one from address so we can use a 32bit load and bswap
     LoadAndSwap(32, scratch1, data);
-    if (cpu_info.bFastBMI2)
+    if (cpu_info.bBMI2FastParallelBitOps)
     {
       MOV(32, R(scratch2), Imm32(0xFCFCFCFC));
       PDEP(32, scratch1, scratch1, R(scratch2));
diff --git a/Source/Core/VideoCommon/VideoConfig.cpp b/Source/Core/VideoCommon/VideoConfig.cpp
index c1ff163f9c..e13e8d96ac 100644
--- a/Source/Core/VideoCommon/VideoConfig.cpp
+++ b/Source/Core/VideoCommon/VideoConfig.cpp
@@ -179,8 +179,8 @@ bool VideoConfig::UsingUberShaders() const
 
 static u32 GetNumAutoShaderCompilerThreads()
 {
-  // Automatic number. We use clamp(cpus - 3, 1, 4).
-  return static_cast<u32>(std::min(std::max(cpu_info.num_cores - 3, 1), 4));
+  // Automatic number.
+  return static_cast<u32>(std::clamp(cpu_info.num_cores - 3, 1, 4));
 }
 
 static u32 GetNumAutoShaderPreCompilerThreads()