From aa844899c937bde5d2b24f276b59997e5b668bde Mon Sep 17 00:00:00 2001 From: Abseil Team Date: Thu, 8 Aug 2019 10:56:58 -0700 Subject: Creation of LTS branch "lts_2019_08_08" - 9ee91d3e430fb33a4590486573792eb0fa146c2d Export of internal Abseil changes by Abseil Team - 8efba58a3b656e9b41fb0471ae6453425a61c520 Export of internal Abseil changes by Abseil Team - b49b8d16b67ec6912899684b732e6367f258cfdb Export of internal Abseil changes by Abseil Team - 67222ffc4c83d918ce8395aa61769eeb77df4c4d Export of internal Abseil changes by Abseil Team - c5c4db4f5191fe5e76cbf68dcc71fb28702f7d2b Export of internal Abseil changes by Abseil Team - 14550beb3b7b97195e483fb74b5efb906395c31e Export of internal Abseil changes. by Abseil Team - 52e88ee56b72cf32bc66534d942c7398ce481331 Export of internal Abseil changes. by Abseil Team - 36d37ab992038f52276ca66b9da80c1cf0f57dc2 Export of internal Abseil changes. by Abseil Team - ad1485c8986246b2ae9105e512738d0e97aec887 Export of internal Abseil changes. by Abseil Team - f3840bc5e33ce4932e35986cf3718450c6f02af2 Export of internal Abseil changes. by Abseil Team - 278b26058c036833a4f7f3047d3f4d9296527f87 Export of internal Abseil changes. by Abseil Team - c6c3c1b498e4ee939b24be59cae29d59c3863be8 Export of internal Abseil changes. by Abseil Team - 44efe96dfca674a17b45ca53fc77fb69f1e29bf4 Export of internal Abseil changes. by Abseil Team - 3c98fcc0461bd2a4b9c149d4748a7373a225cf4b Merge pull request #340 from jtsylve/macos_cxx17_fix by Matt Calabrese <38107210+mattcalabrese-google@users.noreply.github.com> - 74d91756c11bc22f9b0108b94da9326f7f9e376f Export of internal Abseil changes. by Abseil Team - e6b050212c859fbaf67abac76105da10ec348274 Export of internal Abseil changes. by Abseil Team - c964fcffac27bd4a9ff67fe393410dd1146ef8b8 Export of internal Abseil changes. by Abseil Team - 72e09a54d993b192db32be14c65adf7e9bd08c31 Export of internal Abseil changes. by Abseil Team - d65e19dfcd8697076f68598c0131c6930cdcd74d Export of internal Abseil changes. by Abseil Team - 5162fc83d2f3b79a9753ed59594c43966afdd37a Merge pull request #336 from shields/patch-2 by Shaindel Schwartz <31392632+shaindelschwartz@users.noreply.github.com> - 0389f7bf58fa41f35b3ad60be61d32f31e4f8ed6 Merge pull request #335 from shields/patch-1 by Shaindel Schwartz <31392632+shaindelschwartz@users.noreply.github.com> - e9324d926a9189e222741fce6e676f0944661a72 Export of internal Abseil changes. by Abseil Team - 43ef2148c0936ebf7cb4be6b19927a9d9d145b8f Export of internal Abseil changes. by Abseil Team - a13d3df2b3ba68aeead92e2d078fba0510d55024 Merge pull request #323 from gosnik/master by Gennadiy Rozental - 310a11865c97c5cdcc42a4ee2c2e3578423afe69 Merge pull request #324 from RasPat1/patch-1 by Gennadiy Rozental - 8f11724067248acc330b4d1f12f0c76d03f2cfb1 Export of internal Abseil changes. by Abseil Team - b1dd425423380126f6441ce4fbb6f8f6c75b793a Export of internal Abseil changes. by Abseil Team - 361cb8a9db2f2130442389fd80593255be26d681 Export of internal Abseil changes. by Abseil Team - 0238ab0a831f179518c1a814f9584e99da2d75a3 Merge pull request #321 from christoph-cullmann/c4245_fix... by Xiaoyi Zhang - 61c9bf3e3e1c28a4aa6d7f1be4b37fd473bb5529 Export of internal Abseil changes. by Abseil Team - bc9101f9982391019521161a36179b52555ed212 Merge pull request #320 from christoph-cullmann/master by Xiaoyi Zhang - 2f76a9bf50046e396138cc8eeb3cdc17b7a5ac24 Export of internal Abseil changes. by Abseil Team - 4adaf5490921f13028b55018c9f550277de5aebb Export of internal Abseil changes. by Abseil Team - 27c30ec671cb7b5ba84c4e79feff7fd0b0ac6338 Avoid undefined behavior when nullptr is passed to memcpy... by Roman Gershman - ce65f5ac3cbf897bb5e3de1a51d80fd00866abaa Export of internal Abseil changes. by Abseil Team - a18fc7461e7409c2ad64e28537261db1e02e76fa Export of internal Abseil changes. by Abseil Team - 8a394b19c149cab50534b04c5e21d42bc2217a7d Export of internal Abseil changes. by Abseil Team - daf381e8535a1f1f1b8a75966a74e7cca63dee89 Export of internal Abseil changes. by Abseil Team - fa00c321073c7ea40a4fc3dfc8a06309eae3d025 Export of internal Abseil changes. by Abseil Team - 436ba6c4a0ea3a06eca6e055f9c8d296bf3bae12 Export of internal Abseil changes. by Abseil Team - 0cbdc774b97f7e80ab60dbe2ed4eaca3b2e33fc8 Export of internal Abseil changes. by Abseil Team - 27c2f6e2f3b5929fbd322b0f0ca392eb02efd9f8 Export of internal Abseil changes. by Abseil Team - aa468ad75539619b47979911297efbb629c52e44 Export of internal Abseil changes. by Abseil Team - cd86d0d20ab167c33b23d3875db68d1d4bad3a3b Export of internal Abseil changes. by Abseil Team - 33841c5c963aa9c3f096ef8e6c1e71624b941940 Export of internal Abseil changes. by Abseil Team - ca3f87560a0eef716195cadf66dc6b938a579ec6 Export of internal Abseil changes. by Abseil Team - d902eb869bcfacc1bad14933ed9af4bed006d481 Export of internal Abseil changes. by Abseil Team - a02f62f456f2c4a7ecf2be3104fe0c6e16fbad9a Export of internal Abseil changes. by Abseil Team - 0b545b460141b882b244a1efcef7621d59278160 Export of internal Abseil changes. by Abseil Team - dbae8764fbd429bf7d7745e24bcf73962177a7c0 Export of internal Abseil changes. by Abseil Team - 044da8a29c923506af0f0b46bc46f43c1e1300b5 Export of internal Abseil changes. by Abseil Team - 6cc6ac44e065b9e8975fadfd6ccb99cbcf89aac4 Export of internal Abseil changes. by Abseil Team - 666fc1266bccfd8e6eaaa084e7b42580bb8eb199 Export of internal Abseil changes. by Abseil Team - 93dfcf74cb5fccae3da07897d8613ae6cab958a0 Export of internal Abseil changes. by Abseil Team - 2c8421e1c6cef0da9e8a20b01c15256ec9ec116d Export of internal Abseil changes. by Abseil Team - 5b65c4af5107176555b23a638e5947686410ac1f Export of internal Abseil changes. by Abseil Team - eab2078b53c9e3d9d240135c09d27e3393acb50a Export of internal Abseil changes. by Abseil Team - 253eb7416421661873afbaa33828a850db978541 [CMake] Set correct flags for clang-cl (#278) by Loo Rong Jie - e75672f6afc7e8f23ee7b532e86d1b3b9be3984e Export of internal Abseil changes. by Abseil Team - bf29470384a101b307873b26d358433138c857fc Export of internal Abseil changes. by Abseil Team - 6fd827124facd8336981e73218997f9e73029b4f Merge pull request #280 from chiumichael/master by Derek Mauro <761129+derekmauro@users.noreply.github.com> - 7c7754fb3ed9ffb57d35fe8658f3ba4d73a31e72 Export of internal Abseil changes. by Abseil Team - 256be563447a315f2a7993ec669460ba475fa86a Export of internal Abseil changes. by Abseil Team - 88a152ae747c3c42dc9167d46c590929b048d436 Export of internal Abseil changes. by Abseil Team - c1cecb25a94c075725e9d2640f6b978a8f61957b Implement Span::first and Span::last from C++20 (#274) by Girts - 38b704384cd2f17590b3922b97744be0b43622c9 Changed HTTP URLs to HTTPS where possible (#270) by nik7273 - febc5ee6a92d0eb7dac1fceaa6c648cf6521b4dc Export of internal Abseil changes. by Abseil Team - 9fdf5e5b805412cb2a2e624d3e9a11588120465f Export of internal Abseil changes. by Abseil Team - 419f3184f8ebcdb23105295eadd2a569f3351eb9 Export of internal Abseil changes. by Abseil Team - b312c3cb53a0aad75a85ac2bf57c4a614fbd48d4 Export of internal Abseil changes. by Abseil Team - 308ce31528a7edfa39f5f6d36142278a0ae1bf45 Export of internal Abseil changes. by Abseil Team - 93d155bc4414f6c121bb1f19dba9fdb27c8943bc Export of internal Abseil changes. by Abseil Team - 426eaa4aa44e4580418bee46c1bd13911151bfb1 Export of internal Abseil changes. by Abseil Team - 2901ec32a919311384d6ad4194e2d927c06831f7 Export of internal Abseil changes. by Abseil Team - d78310fe5a82f2e0e6e16509ef8079c8d7e4674e Export of internal Abseil changes. by Abseil Team - a4cb1c8ba61531a63f9d309eea01ac1d43d8371d Export of internal Abseil changes. by Abseil Team - 540e2537b92cd4abfae6ceddfe24304345461f32 Export of internal Abseil changes. by Abseil Team - 89ea0c5ff34aaa5855cfc7aa41f323b8a0ef0ede Merge pull request #255 from uilianries/hotfix/conan by ahedberg - 5e0dcf72c64fae912184d2e0de87195fe8f0a425 Export of internal Abseil changes. by Abseil Team - 0dffca4e36791c7beeda04da10460b534283948a Export of internal Abseil changes. by Abseil Team - 6b4201f9ef650637510a21b8d6cbcc3bee4a606f Fix GCC8 warnings by Boris Staletic - 0b1e6d417b414aad9282e32e8c49c719edeb63c1 Export of internal Abseil changes. by Abseil Team - efccc502606bed768e50a6cd5806d8eb13e4e935 Export of internal Abseil changes. by Abseil Team - 5e6a78131f7bd5940218462c07d88cdefdd75dbe Export of internal Abseil changes. by Abseil Team - 5eea0f713c14ac17788b83e496f11903f8e2bbb0 Export of internal Abseil changes. by Abseil Team - 66f9becbb98ecc083f4db349b4b1e0ca9de93b15 Export of internal Abseil changes. by Abseil Team - 018b4db1d73ec8238e6dc4b17fd9e1fd7468d0ed Export of internal Abseil changes. by Abseil Team - 9449ae94397f2fd683851348e25ed8c93f75b3b9 Merge pull request #243 from ThomsonTan/FixIntrinsic by Alex Strelnikov - b16aeb6756bdab08cdf12d40baab5b51f7d15b16 Export of internal Abseil changes. by Abseil Team - 7ffbe09f3d85504bd018783bbe1e2c12992fe47c Export of internal Abseil changes. by Abseil Team - 01b471d9f3ebef27f5aaca14b66509099fa8cd6c Export of internal Abseil changes. by Abseil Team - 7bd8f36c741c7cbe311611d7981bf38ba04c6fef Export of internal Abseil changes. by Abseil Team - 968a34ffdaadd7db062a9621dfbdf8b2d16e05af Export of internal Abseil changes. by Abseil Team - 3e2e9b5557e76d098de4b8a2a659125b98ca519b Merge pull request #231 from uilianries/feature/conan by Mark Barolak - 111ca7060a6ff50115ca85b59f6b5d8c8c5e9105 Export of internal Abseil changes. by Abseil Team - 389ec3f906f018661a5308458d623d01f96d7b23 Export of internal Abseil changes. by Abseil Team - 8fbcdb90952c57828c4a9c2f6d79fcd7cae9088f Export of internal Abseil changes. by Abseil Team - 455dc17ba1af9635f0b60155bc565bc572a1e722 Export of internal Abseil changes. by Abseil Team - f197d7c72a54064cfde5a2058f1513a4a0ee36fb Export of internal Abseil changes. by Abseil Team - 284378a71b32dfb3af4e3661f585e671d1b603a3 Export of internal Abseil changes. by Abseil Team GitOrigin-RevId: 9ee91d3e430fb33a4590486573792eb0fa146c2d Change-Id: Ia06e548bc106cc9d136f6c65714be6645317aced --- absl/random/internal/randen_hwaes.cc | 704 +++++++++++++++++++++++++++++++++++ 1 file changed, 704 insertions(+) create mode 100644 absl/random/internal/randen_hwaes.cc (limited to 'absl/random/internal/randen_hwaes.cc') diff --git a/absl/random/internal/randen_hwaes.cc b/absl/random/internal/randen_hwaes.cc new file mode 100644 index 00000000..d7eed8b2 --- /dev/null +++ b/absl/random/internal/randen_hwaes.cc @@ -0,0 +1,704 @@ +// Copyright 2017 The Abseil Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// https://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// HERMETIC NOTE: The randen_hwaes target must not introduce duplicate +// symbols from arbitrary system and other headers, since it may be built +// with different flags from other targets, using different levels of +// optimization, potentially introducing ODR violations. + +#include "absl/random/internal/randen_hwaes.h" + +#include +#include + +#include "absl/random/internal/platform.h" + +// ABSL_HAVE_ATTRIBUTE +#if !defined(ABSL_HAVE_ATTRIBUTE) +#ifdef __has_attribute +#define ABSL_HAVE_ATTRIBUTE(x) __has_attribute(x) +#else +#define ABSL_HAVE_ATTRIBUTE(x) 0 +#endif +#endif + +#if ABSL_HAVE_ATTRIBUTE(always_inline) || \ + (defined(__GNUC__) && !defined(__clang__)) +#define ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE \ + __attribute__((always_inline)) +#elif defined(_MSC_VER) +// We can achieve something similar to attribute((always_inline)) with MSVC by +// using the __forceinline keyword, however this is not perfect. MSVC is +// much less aggressive about inlining, and even with the __forceinline keyword. +#define ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE __forceinline +#else +#define ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE +#endif + +// ABSL_ATTRIBUTE_FLATTEN enables much more aggressive inlining within +// the indicated function. +#undef ABSL_ATTRIBUTE_FLATTEN +#if ABSL_HAVE_ATTRIBUTE(flatten) || (defined(__GNUC__) && !defined(__clang__)) +#define ABSL_ATTRIBUTE_FLATTEN __attribute__((flatten)) +#else +#define ABSL_ATTRIBUTE_FLATTEN +#endif + +// ABSL_RANDEN_HWAES_IMPL indicates whether this file will contain +// a hardware accelerated implementation of randen, or whether it +// will contain stubs that exit the process. +#if defined(ABSL_ARCH_X86_64) || defined(ABSL_ARCH_X86_32) +// The platform.h directives are sufficient to indicate whether +// we should build accelerated implementations for x86. +#if (ABSL_HAVE_ACCELERATED_AES || ABSL_RANDOM_INTERNAL_AES_DISPATCH) +#define ABSL_RANDEN_HWAES_IMPL 1 +#endif +#elif defined(ABSL_ARCH_PPC) +// The platform.h directives are sufficient to indicate whether +// we should build accelerated implementations for PPC. +// +// NOTE: This has mostly been tested on 64-bit Power variants, +// and not embedded cpus such as powerpc32-8540 +#if ABSL_HAVE_ACCELERATED_AES +#define ABSL_RANDEN_HWAES_IMPL 1 +#endif +#elif defined(ABSL_ARCH_ARM) || defined(ABSL_ARCH_AARCH64) +// ARM is somewhat more complicated. We might support crypto natively... +#if ABSL_HAVE_ACCELERATED_AES || \ + (defined(__ARM_NEON) && defined(__ARM_FEATURE_CRYPTO)) +#define ABSL_RANDEN_HWAES_IMPL 1 + +#elif ABSL_RANDOM_INTERNAL_AES_DISPATCH && !defined(__APPLE__) && \ + (defined(__GNUC__) && __GNUC__ > 4 || __GNUC__ == 4 && __GNUC_MINOR__ > 9) +// ...or, on GCC, we can use an ASM directive to +// instruct the assember to allow crypto instructions. +#define ABSL_RANDEN_HWAES_IMPL 1 +#define ABSL_RANDEN_HWAES_IMPL_CRYPTO_DIRECTIVE 1 +#endif +#else +// HWAES is unsupported by these architectures / platforms: +// __myriad2__ +// __mips__ +// +// Other architectures / platforms are unknown. +// +// See the Abseil documentation on supported macros at: +// https://abseil.io/docs/cpp/platforms/macros +#endif + +#if !defined(ABSL_RANDEN_HWAES_IMPL) +// No accelerated implementation is supported. +// The RandenHwAes functions are stubs that print an error and exit. + +#include +#include + +namespace absl { +inline namespace lts_2019_08_08 { +namespace random_internal { + +// No accelerated implementation. +bool HasRandenHwAesImplementation() { return false; } + +// NOLINTNEXTLINE +const void* RandenHwAes::GetKeys() { + // Attempted to dispatch to an unsupported dispatch target. + const int d = ABSL_RANDOM_INTERNAL_AES_DISPATCH; + fprintf(stderr, "AES Hardware detection failed (%d).\n", d); + exit(1); + return nullptr; +} + +// NOLINTNEXTLINE +void RandenHwAes::Absorb(const void*, void*) { + // Attempted to dispatch to an unsupported dispatch target. + const int d = ABSL_RANDOM_INTERNAL_AES_DISPATCH; + fprintf(stderr, "AES Hardware detection failed (%d).\n", d); + exit(1); +} + +// NOLINTNEXTLINE +void RandenHwAes::Generate(const void*, void*) { + // Attempted to dispatch to an unsupported dispatch target. + const int d = ABSL_RANDOM_INTERNAL_AES_DISPATCH; + fprintf(stderr, "AES Hardware detection failed (%d).\n", d); + exit(1); +} + +} // namespace random_internal +} // inline namespace lts_2019_08_08 +} // namespace absl + +#else // defined(ABSL_RANDEN_HWAES_IMPL) +// +// Accelerated implementations are supported. +// We need the per-architecture includes and defines. +// + +#include "absl/random/internal/randen_traits.h" + +// ABSL_FUNCTION_ALIGN32 defines a 32-byte alignment attribute +// for the functions in this file. +// +// NOTE: Determine whether we actually have any wins from ALIGN32 +// using microbenchmarks. If not, remove. +#undef ABSL_FUNCTION_ALIGN32 +#if ABSL_HAVE_ATTRIBUTE(aligned) || (defined(__GNUC__) && !defined(__clang__)) +#define ABSL_FUNCTION_ALIGN32 __attribute__((aligned(32))) +#else +#define ABSL_FUNCTION_ALIGN32 +#endif + +// TARGET_CRYPTO defines a crypto attribute for each architecture. +// +// NOTE: Evaluate whether we should eliminate ABSL_TARGET_CRYPTO. +#if (defined(__clang__) || defined(__GNUC__)) +#if defined(ABSL_ARCH_X86_64) || defined(ABSL_ARCH_X86_32) +#define ABSL_TARGET_CRYPTO __attribute__((target("aes"))) +#elif defined(ABSL_ARCH_PPC) +#define ABSL_TARGET_CRYPTO __attribute__((target("crypto"))) +#else +#define ABSL_TARGET_CRYPTO +#endif +#else +#define ABSL_TARGET_CRYPTO +#endif + +#if defined(ABSL_ARCH_PPC) +// NOTE: Keep in mind that PPC can operate in little-endian or big-endian mode, +// however the PPC altivec vector registers (and thus the AES instructions) +// always operate in big-endian mode. + +#include +// #defines vector __vector; in C++, this is bad form. +#undef vector + +// Rely on the PowerPC AltiVec vector operations for accelerated AES +// instructions. GCC support of the PPC vector types is described in: +// https://gcc.gnu.org/onlinedocs/gcc-4.9.0/gcc/PowerPC-AltiVec_002fVSX-Built-in-Functions.html +// +// Already provides operator^=. +using Vector128 = __vector unsigned long long; // NOLINT(runtime/int) + +namespace { + +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128 +ReverseBytes(const Vector128& v) { + // Reverses the bytes of the vector. + const __vector unsigned char perm = {15, 14, 13, 12, 11, 10, 9, 8, + 7, 6, 5, 4, 3, 2, 1, 0}; + return vec_perm(v, v, perm); +} + +// WARNING: these load/store in native byte order. It is OK to load and then +// store an unchanged vector, but interpreting the bits as a number or input +// to AES will have undefined results. +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128 +Vector128Load(const void* ABSL_RANDOM_INTERNAL_RESTRICT from) { + return vec_vsx_ld(0, reinterpret_cast(from)); +} + +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void +Vector128Store(const Vector128& v, void* ABSL_RANDOM_INTERNAL_RESTRICT to) { + vec_vsx_st(v, 0, reinterpret_cast(to)); +} + +// One round of AES. "round_key" is a public constant for breaking the +// symmetry of AES (ensures previously equal columns differ afterwards). +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128 +AesRound(const Vector128& state, const Vector128& round_key) { + return Vector128(__builtin_crypto_vcipher(state, round_key)); +} + +// Enables native loads in the round loop by pre-swapping. +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void +SwapEndian(uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT state) { + using absl::random_internal::RandenTraits; + constexpr size_t kLanes = 2; + constexpr size_t kFeistelBlocks = RandenTraits::kFeistelBlocks; + + for (uint32_t branch = 0; branch < kFeistelBlocks; ++branch) { + const Vector128 v = ReverseBytes(Vector128Load(state + kLanes * branch)); + Vector128Store(v, state + kLanes * branch); + } +} + +} // namespace + +#elif defined(ABSL_ARCH_ARM) || defined(ABSL_ARCH_AARCH64) + +// This asm directive will cause the file to be compiled with crypto extensions +// whether or not the cpu-architecture supports it. +#if ABSL_RANDEN_HWAES_IMPL_CRYPTO_DIRECTIVE +asm(".arch_extension crypto\n"); + +// Override missing defines. +#if !defined(__ARM_NEON) +#define __ARM_NEON 1 +#endif + +#if !defined(__ARM_FEATURE_CRYPTO) +#define __ARM_FEATURE_CRYPTO 1 +#endif + +#endif + +// Rely on the ARM NEON+Crypto advanced simd types, defined in . +// uint8x16_t is the user alias for underlying __simd128_uint8_t type. +// http://infocenter.arm.com/help/topic/com.arm.doc.ihi0073a/IHI0073A_arm_neon_intrinsics_ref.pdf +// +// defines the following +// +// typedef __attribute__((neon_vector_type(16))) uint8_t uint8x16_t; +// typedef __attribute__((neon_vector_type(16))) int8_t int8x16_t; +// typedef __attribute__((neon_polyvector_type(16))) int8_t poly8x16_t; +// +// vld1q_v +// vst1q_v +// vaeseq_v +// vaesmcq_v +#include + +// Already provides operator^=. +using Vector128 = uint8x16_t; + +namespace { + +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128 +Vector128Load(const void* ABSL_RANDOM_INTERNAL_RESTRICT from) { + return vld1q_u8(reinterpret_cast(from)); +} + +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void +Vector128Store(const Vector128& v, void* ABSL_RANDOM_INTERNAL_RESTRICT to) { + vst1q_u8(reinterpret_cast(to), v); +} + +// One round of AES. "round_key" is a public constant for breaking the +// symmetry of AES (ensures previously equal columns differ afterwards). +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128 +AesRound(const Vector128& state, const Vector128& round_key) { + // It is important to always use the full round function - omitting the + // final MixColumns reduces security [https://eprint.iacr.org/2010/041.pdf] + // and does not help because we never decrypt. + // + // Note that ARM divides AES instructions differently than x86 / PPC, + // And we need to skip the first AddRoundKey step and add an extra + // AddRoundKey step to the end. Lucky for us this is just XOR. + return vaesmcq_u8(vaeseq_u8(state, uint8x16_t{})) ^ round_key; +} + +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void +SwapEndian(uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT) {} + +} // namespace + +#elif defined(ABSL_ARCH_X86_64) || defined(ABSL_ARCH_X86_32) +// On x86 we rely on the aesni instructions +#include + +namespace { + +// Vector128 class is only wrapper for __m128i, benchmark indicates that it's +// faster than using __m128i directly. +class Vector128 { + public: + // Convert from/to intrinsics. + inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE explicit Vector128( + const __m128i& Vector128) + : data_(Vector128) {} + + inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE __m128i data() const { + return data_; + } + + inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128& operator^=( + const Vector128& other) { + data_ = _mm_xor_si128(data_, other.data()); + return *this; + } + + private: + __m128i data_; +}; + +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128 +Vector128Load(const void* ABSL_RANDOM_INTERNAL_RESTRICT from) { + return Vector128(_mm_load_si128(reinterpret_cast(from))); +} + +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void +Vector128Store(const Vector128& v, void* ABSL_RANDOM_INTERNAL_RESTRICT to) { + _mm_store_si128(reinterpret_cast<__m128i * ABSL_RANDOM_INTERNAL_RESTRICT>(to), + v.data()); +} + +// One round of AES. "round_key" is a public constant for breaking the +// symmetry of AES (ensures previously equal columns differ afterwards). +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128 +AesRound(const Vector128& state, const Vector128& round_key) { + // It is important to always use the full round function - omitting the + // final MixColumns reduces security [https://eprint.iacr.org/2010/041.pdf] + // and does not help because we never decrypt. + return Vector128(_mm_aesenc_si128(state.data(), round_key.data())); +} + +inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void +SwapEndian(uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT) {} + +} // namespace + +#endif + +namespace { + +// u64x2 is a 128-bit, (2 x uint64_t lanes) struct used to store +// the randen_keys. +struct alignas(16) u64x2 { + constexpr u64x2(uint64_t hi, uint64_t lo) +#if defined(ABSL_ARCH_PPC) + // This has been tested with PPC running in little-endian mode; + // We byte-swap the u64x2 structure from little-endian to big-endian + // because altivec always runs in big-endian mode. + : v{__builtin_bswap64(hi), __builtin_bswap64(lo)} { +#else + : v{lo, hi} { +#endif + } + + constexpr bool operator==(const u64x2& other) const { + return v[0] == other.v[0] && v[1] == other.v[1]; + } + + constexpr bool operator!=(const u64x2& other) const { + return !(*this == other); + } + + uint64_t v[2]; +}; // namespace + +#ifdef __clang__ +#pragma clang diagnostic push +#pragma clang diagnostic ignored "-Wunknown-pragmas" +#endif + +// At this point, all of the platform-specific features have been defined / +// implemented. +// +// REQUIRES: using u64x2 = ... +// REQUIRES: using Vector128 = ... +// REQUIRES: Vector128 Vector128Load(void*) {...} +// REQUIRES: void Vector128Store(Vector128, void*) {...} +// REQUIRES: Vector128 AesRound(Vector128, Vector128) {...} +// REQUIRES: void SwapEndian(uint64_t*) {...} +// +// PROVIDES: absl::random_internal::RandenHwAes::Absorb +// PROVIDES: absl::random_internal::RandenHwAes::Generate + +// RANDen = RANDom generator or beetroots in Swiss German. +// 'Strong' (well-distributed, unpredictable, backtracking-resistant) random +// generator, faster in some benchmarks than std::mt19937_64 and pcg64_c32. +// +// High-level summary: +// 1) Reverie (see "A Robust and Sponge-Like PRNG with Improved Efficiency") is +// a sponge-like random generator that requires a cryptographic permutation. +// It improves upon "Provably Robust Sponge-Based PRNGs and KDFs" by +// achieving backtracking resistance with only one Permute() per buffer. +// +// 2) "Simpira v2: A Family of Efficient Permutations Using the AES Round +// Function" constructs up to 1024-bit permutations using an improved +// Generalized Feistel network with 2-round AES-128 functions. This Feistel +// block shuffle achieves diffusion faster and is less vulnerable to +// sliced-biclique attacks than the Type-2 cyclic shuffle. +// +// 3) "Improving the Generalized Feistel" and "New criterion for diffusion +// property" extends the same kind of improved Feistel block shuffle to 16 +// branches, which enables a 2048-bit permutation. +// +// We combine these three ideas and also change Simpira's subround keys from +// structured/low-entropy counters to digits of Pi. + +// Randen constants. +using absl::random_internal::RandenTraits; +constexpr size_t kStateBytes = RandenTraits::kStateBytes; +constexpr size_t kCapacityBytes = RandenTraits::kCapacityBytes; +constexpr size_t kFeistelBlocks = RandenTraits::kFeistelBlocks; +constexpr size_t kFeistelRounds = RandenTraits::kFeistelRounds; +constexpr size_t kFeistelFunctions = RandenTraits::kFeistelFunctions; + +// Independent keys (272 = 2.1 KiB) for the first AES subround of each function. +constexpr size_t kKeys = kFeistelRounds * kFeistelFunctions; + +// INCLUDE keys. +#include "absl/random/internal/randen-keys.inc" + +static_assert(kKeys == kRoundKeys, "kKeys and kRoundKeys must be equal"); +static_assert(round_keys[kKeys - 1] != u64x2(0, 0), + "Too few round_keys initializers"); + +// Number of uint64_t lanes per 128-bit vector; +constexpr size_t kLanes = 2; + +// Block shuffles applies a shuffle to the entire state between AES rounds. +// Improved odd-even shuffle from "New criterion for diffusion property". +inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE ABSL_TARGET_CRYPTO void +BlockShuffle(uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT state) { + static_assert(kFeistelBlocks == 16, "Expecting 16 FeistelBlocks."); + + constexpr size_t shuffle[kFeistelBlocks] = {7, 2, 13, 4, 11, 8, 3, 6, + 15, 0, 9, 10, 1, 14, 5, 12}; + + // The fully unrolled loop without the memcpy improves the speed by about + // 30% over the equivalent loop. + const Vector128 v0 = Vector128Load(state + kLanes * shuffle[0]); + const Vector128 v1 = Vector128Load(state + kLanes * shuffle[1]); + const Vector128 v2 = Vector128Load(state + kLanes * shuffle[2]); + const Vector128 v3 = Vector128Load(state + kLanes * shuffle[3]); + const Vector128 v4 = Vector128Load(state + kLanes * shuffle[4]); + const Vector128 v5 = Vector128Load(state + kLanes * shuffle[5]); + const Vector128 v6 = Vector128Load(state + kLanes * shuffle[6]); + const Vector128 v7 = Vector128Load(state + kLanes * shuffle[7]); + const Vector128 w0 = Vector128Load(state + kLanes * shuffle[8]); + const Vector128 w1 = Vector128Load(state + kLanes * shuffle[9]); + const Vector128 w2 = Vector128Load(state + kLanes * shuffle[10]); + const Vector128 w3 = Vector128Load(state + kLanes * shuffle[11]); + const Vector128 w4 = Vector128Load(state + kLanes * shuffle[12]); + const Vector128 w5 = Vector128Load(state + kLanes * shuffle[13]); + const Vector128 w6 = Vector128Load(state + kLanes * shuffle[14]); + const Vector128 w7 = Vector128Load(state + kLanes * shuffle[15]); + + Vector128Store(v0, state + kLanes * 0); + Vector128Store(v1, state + kLanes * 1); + Vector128Store(v2, state + kLanes * 2); + Vector128Store(v3, state + kLanes * 3); + Vector128Store(v4, state + kLanes * 4); + Vector128Store(v5, state + kLanes * 5); + Vector128Store(v6, state + kLanes * 6); + Vector128Store(v7, state + kLanes * 7); + Vector128Store(w0, state + kLanes * 8); + Vector128Store(w1, state + kLanes * 9); + Vector128Store(w2, state + kLanes * 10); + Vector128Store(w3, state + kLanes * 11); + Vector128Store(w4, state + kLanes * 12); + Vector128Store(w5, state + kLanes * 13); + Vector128Store(w6, state + kLanes * 14); + Vector128Store(w7, state + kLanes * 15); +} + +// Feistel round function using two AES subrounds. Very similar to F() +// from Simpira v2, but with independent subround keys. Uses 17 AES rounds +// per 16 bytes (vs. 10 for AES-CTR). Computing eight round functions in +// parallel hides the 7-cycle AESNI latency on HSW. Note that the Feistel +// XORs are 'free' (included in the second AES instruction). +inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE ABSL_TARGET_CRYPTO const + u64x2* + FeistelRound(uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT state, + const u64x2* ABSL_RANDOM_INTERNAL_RESTRICT keys) { + static_assert(kFeistelBlocks == 16, "Expecting 16 FeistelBlocks."); + + // MSVC does a horrible job at unrolling loops. + // So we unroll the loop by hand to improve the performance. + const Vector128 s0 = Vector128Load(state + kLanes * 0); + const Vector128 s1 = Vector128Load(state + kLanes * 1); + const Vector128 s2 = Vector128Load(state + kLanes * 2); + const Vector128 s3 = Vector128Load(state + kLanes * 3); + const Vector128 s4 = Vector128Load(state + kLanes * 4); + const Vector128 s5 = Vector128Load(state + kLanes * 5); + const Vector128 s6 = Vector128Load(state + kLanes * 6); + const Vector128 s7 = Vector128Load(state + kLanes * 7); + const Vector128 s8 = Vector128Load(state + kLanes * 8); + const Vector128 s9 = Vector128Load(state + kLanes * 9); + const Vector128 s10 = Vector128Load(state + kLanes * 10); + const Vector128 s11 = Vector128Load(state + kLanes * 11); + const Vector128 s12 = Vector128Load(state + kLanes * 12); + const Vector128 s13 = Vector128Load(state + kLanes * 13); + const Vector128 s14 = Vector128Load(state + kLanes * 14); + const Vector128 s15 = Vector128Load(state + kLanes * 15); + + // Encode even blocks with keys. + const Vector128 e0 = AesRound(s0, Vector128Load(keys + 0)); + const Vector128 e2 = AesRound(s2, Vector128Load(keys + 1)); + const Vector128 e4 = AesRound(s4, Vector128Load(keys + 2)); + const Vector128 e6 = AesRound(s6, Vector128Load(keys + 3)); + const Vector128 e8 = AesRound(s8, Vector128Load(keys + 4)); + const Vector128 e10 = AesRound(s10, Vector128Load(keys + 5)); + const Vector128 e12 = AesRound(s12, Vector128Load(keys + 6)); + const Vector128 e14 = AesRound(s14, Vector128Load(keys + 7)); + + // Encode odd blocks with even output from above. + const Vector128 o1 = AesRound(e0, s1); + const Vector128 o3 = AesRound(e2, s3); + const Vector128 o5 = AesRound(e4, s5); + const Vector128 o7 = AesRound(e6, s7); + const Vector128 o9 = AesRound(e8, s9); + const Vector128 o11 = AesRound(e10, s11); + const Vector128 o13 = AesRound(e12, s13); + const Vector128 o15 = AesRound(e14, s15); + + // Store odd blocks. (These will be shuffled later). + Vector128Store(o1, state + kLanes * 1); + Vector128Store(o3, state + kLanes * 3); + Vector128Store(o5, state + kLanes * 5); + Vector128Store(o7, state + kLanes * 7); + Vector128Store(o9, state + kLanes * 9); + Vector128Store(o11, state + kLanes * 11); + Vector128Store(o13, state + kLanes * 13); + Vector128Store(o15, state + kLanes * 15); + + return keys + 8; +} + +// Cryptographic permutation based via type-2 Generalized Feistel Network. +// Indistinguishable from ideal by chosen-ciphertext adversaries using less than +// 2^64 queries if the round function is a PRF. This is similar to the b=8 case +// of Simpira v2, but more efficient than its generic construction for b=16. +inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE ABSL_TARGET_CRYPTO void +Permute(const void* ABSL_RANDOM_INTERNAL_RESTRICT keys, + uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT state) { + const u64x2* ABSL_RANDOM_INTERNAL_RESTRICT keys128 = + static_cast(keys); + + // (Successfully unrolled; the first iteration jumps into the second half) +#ifdef __clang__ +#pragma clang loop unroll_count(2) +#endif + for (size_t round = 0; round < kFeistelRounds; ++round) { + keys128 = FeistelRound(state, keys128); + BlockShuffle(state); + } +} + +} // namespace + +namespace absl { +inline namespace lts_2019_08_08 { +namespace random_internal { + +bool HasRandenHwAesImplementation() { return true; } + +const void* ABSL_TARGET_CRYPTO ABSL_FUNCTION_ALIGN32 ABSL_ATTRIBUTE_FLATTEN +RandenHwAes::GetKeys() { + // Round keys for one AES per Feistel round and branch. + // The canonical implementation uses first digits of Pi. + return round_keys; +} + +// NOLINTNEXTLINE +void ABSL_TARGET_CRYPTO ABSL_FUNCTION_ALIGN32 ABSL_ATTRIBUTE_FLATTEN +RandenHwAes::Absorb(const void* seed_void, void* state_void) { + uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT state = + reinterpret_cast(state_void); + const uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT seed = + reinterpret_cast(seed_void); + + constexpr size_t kCapacityBlocks = kCapacityBytes / sizeof(Vector128); + constexpr size_t kStateBlocks = kStateBytes / sizeof(Vector128); + + static_assert(kCapacityBlocks * sizeof(Vector128) == kCapacityBytes, + "Not i*V"); + static_assert(kCapacityBlocks == 1, "Unexpected Randen kCapacityBlocks"); + static_assert(kStateBlocks == 16, "Unexpected Randen kStateBlocks"); + + Vector128 b1 = Vector128Load(state + kLanes * 1); + b1 ^= Vector128Load(seed + kLanes * 0); + Vector128Store(b1, state + kLanes * 1); + + Vector128 b2 = Vector128Load(state + kLanes * 2); + b2 ^= Vector128Load(seed + kLanes * 1); + Vector128Store(b2, state + kLanes * 2); + + Vector128 b3 = Vector128Load(state + kLanes * 3); + b3 ^= Vector128Load(seed + kLanes * 2); + Vector128Store(b3, state + kLanes * 3); + + Vector128 b4 = Vector128Load(state + kLanes * 4); + b4 ^= Vector128Load(seed + kLanes * 3); + Vector128Store(b4, state + kLanes * 4); + + Vector128 b5 = Vector128Load(state + kLanes * 5); + b5 ^= Vector128Load(seed + kLanes * 4); + Vector128Store(b5, state + kLanes * 5); + + Vector128 b6 = Vector128Load(state + kLanes * 6); + b6 ^= Vector128Load(seed + kLanes * 5); + Vector128Store(b6, state + kLanes * 6); + + Vector128 b7 = Vector128Load(state + kLanes * 7); + b7 ^= Vector128Load(seed + kLanes * 6); + Vector128Store(b7, state + kLanes * 7); + + Vector128 b8 = Vector128Load(state + kLanes * 8); + b8 ^= Vector128Load(seed + kLanes * 7); + Vector128Store(b8, state + kLanes * 8); + + Vector128 b9 = Vector128Load(state + kLanes * 9); + b9 ^= Vector128Load(seed + kLanes * 8); + Vector128Store(b9, state + kLanes * 9); + + Vector128 b10 = Vector128Load(state + kLanes * 10); + b10 ^= Vector128Load(seed + kLanes * 9); + Vector128Store(b10, state + kLanes * 10); + + Vector128 b11 = Vector128Load(state + kLanes * 11); + b11 ^= Vector128Load(seed + kLanes * 10); + Vector128Store(b11, state + kLanes * 11); + + Vector128 b12 = Vector128Load(state + kLanes * 12); + b12 ^= Vector128Load(seed + kLanes * 11); + Vector128Store(b12, state + kLanes * 12); + + Vector128 b13 = Vector128Load(state + kLanes * 13); + b13 ^= Vector128Load(seed + kLanes * 12); + Vector128Store(b13, state + kLanes * 13); + + Vector128 b14 = Vector128Load(state + kLanes * 14); + b14 ^= Vector128Load(seed + kLanes * 13); + Vector128Store(b14, state + kLanes * 14); + + Vector128 b15 = Vector128Load(state + kLanes * 15); + b15 ^= Vector128Load(seed + kLanes * 14); + Vector128Store(b15, state + kLanes * 15); +} + +// NOLINTNEXTLINE +void ABSL_TARGET_CRYPTO ABSL_FUNCTION_ALIGN32 ABSL_ATTRIBUTE_FLATTEN +RandenHwAes::Generate(const void* keys, void* state_void) { + static_assert(kCapacityBytes == sizeof(Vector128), "Capacity mismatch"); + + uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT state = + reinterpret_cast(state_void); + + const Vector128 prev_inner = Vector128Load(state); + + SwapEndian(state); + + Permute(keys, state); + + SwapEndian(state); + + // Ensure backtracking resistance. + Vector128 inner = Vector128Load(state); + inner ^= prev_inner; + Vector128Store(inner, state); +} + +#ifdef __clang__ +#pragma clang diagnostic pop +#endif + +} // namespace random_internal +} // inline namespace lts_2019_08_08 +} // namespace absl + +#endif // (ABSL_RANDEN_HWAES_IMPL) -- cgit v1.2.3