Random: Support runtime dispatch on AArch64 macOS

This lets us dynamically determine if we should attempt to use HW
instructions for AES.

PiperOrigin-RevId: 716350656
Change-Id: I1ead17b7b44373850b3591307b7f6d3f9147d5f9

Author: majnemer

absl/random/CMakeLists.txt

@@ -853,6 +853,7 @@ absl_cc_library(
     absl::random_internal_platform
     absl::random_internal_randen_hwaes_impl
     absl::config
+    absl::optional
 )
 
 # Internal-only target, do not depend on directly.

absl/random/internal/BUILD.bazel

@@ -403,6 +403,7 @@ cc_library(
         ":platform",
         ":randen_hwaes_impl",
         "//absl/base:config",
+        "//absl/types:optional",
     ],
 )
 

absl/random/internal/platform.h

@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+// SKIP_ABSL_INLINE_NAMESPACE_CHECK
+
 #ifndef ABSL_RANDOM_INTERNAL_PLATFORM_H_
 #define ABSL_RANDOM_INTERNAL_PLATFORM_H_
 
@@ -134,16 +136,23 @@
 // accelerated Randen implementation.
 #define ABSL_RANDOM_INTERNAL_AES_DISPATCH 0
 
-#if defined(ABSL_ARCH_X86_64)
+// iOS does not support dispatch, even on x86, since applications
+// should be bundled as fat binaries, with a different build tailored for
+// each specific supported platform/architecture.
+#if (defined(TARGET_OS_IPHONE) && TARGET_OS_IPHONE) || \
+    (defined(TARGET_OS_IPHONE_SIMULATOR) && TARGET_OS_IPHONE_SIMULATOR)
+#undef ABSL_RANDOM_INTERNAL_AES_DISPATCH
+#define ABSL_RANDOM_INTERNAL_AES_DISPATCH 0
+#elif defined(ABSL_ARCH_X86_64)
 // Dispatch is available on x86_64
 #undef ABSL_RANDOM_INTERNAL_AES_DISPATCH
 #define ABSL_RANDOM_INTERNAL_AES_DISPATCH 1
 #elif defined(__linux__) && defined(ABSL_ARCH_PPC)
 // Or when running linux PPC
 #undef ABSL_RANDOM_INTERNAL_AES_DISPATCH
 #define ABSL_RANDOM_INTERNAL_AES_DISPATCH 1
-#elif defined(__linux__) && defined(ABSL_ARCH_AARCH64)
-// Or when running linux AArch64
+#elif (defined(__linux__) || defined(__APPLE__)) && defined(ABSL_ARCH_AARCH64)
+// Or when running linux or macOS AArch64
 #undef ABSL_RANDOM_INTERNAL_AES_DISPATCH
 #define ABSL_RANDOM_INTERNAL_AES_DISPATCH 1
 #elif defined(__linux__) && defined(ABSL_ARCH_ARM) && (__ARM_ARCH >= 8)
@@ -159,13 +168,4 @@
 #define ABSL_RANDOM_INTERNAL_AES_DISPATCH 0
 #endif
 
-// iOS does not support dispatch, even on x86, since applications
-// should be bundled as fat binaries, with a different build tailored for
-// each specific supported platform/architecture.
-#if (defined(TARGET_OS_IPHONE) && TARGET_OS_IPHONE) || \
-    (defined(TARGET_OS_IPHONE_SIMULATOR) && TARGET_OS_IPHONE_SIMULATOR)
-#undef ABSL_RANDOM_INTERNAL_AES_DISPATCH
-#define ABSL_RANDOM_INTERNAL_AES_DISPATCH 0
-#endif
-
 #endif  // ABSL_RANDOM_INTERNAL_PLATFORM_H_

absl/random/internal/randen_detect.cc

@@ -19,10 +19,21 @@
 
 #include "absl/random/internal/randen_detect.h"
 
+#if defined(__APPLE__) && defined(__aarch64__)
+#if defined(__has_include)
+#if __has_include(<arm/cpu_capabilities_public.h>)
+#include <arm/cpu_capabilities_public.h>
+#endif
+#endif
+#include <sys/sysctl.h>
+#include <sys/types.h>
+#endif
+
 #include <cstdint>
 #include <cstring>
 
 #include "absl/random/internal/platform.h"
+#include "absl/types/optional.h"  // IWYU pragma: keep
 
 #if !defined(__UCLIBC__) && defined(__GLIBC__) && \
     (__GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ >= 16))
@@ -102,6 +113,19 @@ static uint32_t GetAuxval(uint32_t hwcap_type) {
 
 #endif
 
+#if defined(__APPLE__) && defined(ABSL_ARCH_AARCH64)
+template <typename T>
+static absl::optional<T> ReadSysctlByName(const char* name) {
+  T val;
+  size_t val_size = sizeof(T);
+  int ret = sysctlbyname(name, &val, &val_size, nullptr, 0);
+  if (ret == -1) {
+    return std::nullopt;
+  }
+  return val;
+}
+#endif
+
 namespace absl {
 ABSL_NAMESPACE_BEGIN
 namespace random_internal {
@@ -122,14 +146,18 @@ namespace random_internal {
 //
 // Fon non-x86 it is much more complicated.
 //
-// 2. When ABSL_INTERNAL_USE_GETAUXVAL is defined, use getauxval() (either
+// 2. Try to use __builtin_cpu_supports.
+//
+// 3. When ABSL_INTERNAL_USE_GETAUXVAL is defined, use getauxval() (either
 //    the direct c-library version, or the android probing version which loads
 //    libc), and read the hardware capability bits.
 //    This is based on the technique used by boringssl uses to detect
 //    cpu capabilities, and should allow us to enable crypto in the android
 //    builds where it is supported.
 //
-// 3. Use the default for the compiler architecture.
+// 4. When __APPLE__ is defined on AARCH64, use sysctlbyname().
+//
+// 5. Use the default for the compiler architecture.
 //
 
 bool CPUSupportsRandenHwAes() {
@@ -139,8 +167,14 @@ bool CPUSupportsRandenHwAes() {
   __cpuid(reinterpret_cast<int*>(regs), 1);
   return regs[2] & (1 << 25);  // AES
 
+#elif defined(ABSL_ARCH_AARCH64) && ABSL_HAVE_BUILTIN(__builtin_cpu_supports)
+  // For AARCH64: Require crypto+neon
+  // http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0500f/CIHBIBBA.html
+  //
+  // __builtin_cpu_supports spells "crypt" as "aes" and "neon" as "simd".
+  return __builtin_cpu_supports("aes+simd");
 #elif defined(ABSL_INTERNAL_USE_GETAUXVAL)
-  // 2. Use getauxval() to read the hardware bits and determine
+  // 3. Use getauxval() to read the hardware bits and determine
   // cpu capabilities.
 
 #define AT_HWCAP 16
@@ -178,8 +212,36 @@ bool CPUSupportsRandenHwAes() {
   return ((hwcap & kNEON) != 0) && ((hwcap & kAES) != 0);
 #endif
 
+#elif defined(__APPLE__) && defined(ABSL_ARCH_AARCH64)
+  // 4. Use sysctlbyname.
+
+  // Newer XNU kernels support querying all capabilities in a single
+  // sysctlbyname.
+#if defined(CAP_BIT_AdvSIMD) && defined(CAP_BIT_FEAT_AES)
+  static const absl::optional<uint64_t> caps =
+      ReadSysctlByName<uint64_t>("hw.optional.arm.caps");
+  if (caps.has_value()) {
+    constexpr uint64_t kNeonAndAesCaps =
+        (uint64_t{1} << CAP_BIT_AdvSIMD) | (uint64_t{1} << CAP_BIT_FEAT_AES);
+    return (*caps & kNeonAndAesCaps) == kNeonAndAesCaps;
+  }
+#endif
+
+  // https://developer.apple.com/documentation/kernel/1387446-sysctlbyname/determining_instruction_set_characteristics#overview
+  static const absl::optional<int> adv_simd =
+      ReadSysctlByName<int>("hw.optional.AdvSIMD");
+  if (adv_simd.value_or(0) == 0) {
+    return false;
+  }
+  // https://developer.apple.com/documentation/kernel/1387446-sysctlbyname/determining_instruction_set_characteristics#3918855
+  static const absl::optional<int> feat_aes =
+      ReadSysctlByName<int>("hw.optional.arm.FEAT_AES");
+  if (feat_aes.value_or(0) == 0) {
+    return false;
+  }
+  return true;
 #else  // ABSL_INTERNAL_USE_GETAUXVAL
-  // 3. By default, assume that the compiler default.
+  // 5. By default, assume that the compiler default.
   return ABSL_HAVE_ACCELERATED_AES ? true : false;
 
 #endif
@@ -215,9 +277,6 @@ bool CPUSupportsRandenHwAes() {
   //   __asm __volatile("mrs %0, id_aa64isar0_el1" :"=&r" (val));
   //
   // * Use a CPUID-style heuristic database.
-  //
-  // * On Apple (__APPLE__), AES is available on Arm v8.
-  //   https://stackoverflow.com/questions/45637888/how-to-determine-armv8-features-at-runtime-on-ios
 }
 
 #if defined(__clang__)

absl/random/internal/randen_hwaes_test.cc

@@ -89,7 +89,7 @@ int main(int argc, char* argv[]) {
   LOG(INFO) << "HasRandenHwAesImplementation = " << x;
 
   int y = absl::random_internal::CPUSupportsRandenHwAes();
-  LOG(INFO) << "CPUSupportsRandenHwAes = " << x;
+  LOG(INFO) << "CPUSupportsRandenHwAes = " << y;
 
   if (!x || !y) {
     LOG(INFO) << "Skipping Randen HWAES tests.";