aws-crt 0.1.9 → 0.2.0
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +5 -0
- data/VERSION +1 -1
- data/aws-crt-ffi/crt/aws-c-auth/include/aws/auth/auth.h +1 -0
- data/aws-crt-ffi/crt/aws-c-auth/include/aws/auth/aws_imds_client.h +5 -0
- data/aws-crt-ffi/crt/aws-c-auth/include/aws/auth/credentials.h +5 -0
- data/aws-crt-ffi/crt/aws-c-auth/include/aws/auth/private/aws_signing.h +1 -0
- data/aws-crt-ffi/crt/aws-c-auth/include/aws/auth/private/credentials_utils.h +2 -0
- data/aws-crt-ffi/crt/aws-c-auth/include/aws/auth/signing_config.h +1 -0
- data/aws-crt-ffi/crt/aws-c-auth/source/auth.c +3 -1
- data/aws-crt-ffi/crt/aws-c-auth/source/aws_imds_client.c +146 -63
- data/aws-crt-ffi/crt/aws-c-auth/source/aws_signing.c +41 -19
- data/aws-crt-ffi/crt/aws-c-auth/source/credentials_provider_imds.c +1 -0
- data/aws-crt-ffi/crt/aws-c-auth/source/credentials_utils.c +1 -0
- data/aws-crt-ffi/crt/aws-c-auth/source/signable_http_request.c +2 -1
- data/aws-crt-ffi/crt/aws-c-auth/source/signing_config.c +25 -0
- data/aws-crt-ffi/crt/aws-c-auth/tests/CMakeLists.txt +3 -0
- data/aws-crt-ffi/crt/aws-c-auth/tests/aws_imds_client_test.c +197 -31
- data/aws-crt-ffi/crt/aws-c-auth/tests/credentials_provider_imds_tests.c +16 -18
- data/aws-crt-ffi/crt/aws-c-auth/tests/sigv4_signing_tests.c +3 -1
- data/aws-crt-ffi/crt/aws-c-cal/include/aws/cal/private/opensslcrypto_common.h +22 -0
- data/aws-crt-ffi/crt/aws-c-cal/source/darwin/commoncrypto_aes.c +46 -17
- data/aws-crt-ffi/crt/aws-c-cal/source/unix/openssl_aes.c +1 -0
- data/aws-crt-ffi/crt/aws-c-cal/source/unix/openssl_platform_init.c +7 -0
- data/aws-crt-ffi/crt/aws-c-cal/source/unix/openssl_rsa.c +59 -2
- data/aws-crt-ffi/crt/aws-c-cal/source/unix/opensslcrypto_ecc.c +1 -0
- data/aws-crt-ffi/crt/aws-c-common/CMakeLists.txt +13 -1
- data/aws-crt-ffi/crt/aws-c-common/THIRD-PARTY-LICENSES.txt +28 -7
- data/aws-crt-ffi/crt/aws-c-common/bin/system_info/CMakeLists.txt +18 -0
- data/aws-crt-ffi/crt/aws-c-common/bin/system_info/print_system_info.c +48 -0
- data/aws-crt-ffi/crt/aws-c-common/include/aws/common/allocator.h +23 -0
- data/aws-crt-ffi/crt/aws-c-common/include/aws/common/byte_buf.h +12 -0
- data/aws-crt-ffi/crt/aws-c-common/include/aws/common/cross_process_lock.h +35 -0
- data/aws-crt-ffi/crt/aws-c-common/include/aws/common/hash_table.h +1 -0
- data/aws-crt-ffi/crt/aws-c-common/include/aws/common/priority_queue.h +24 -0
- data/aws-crt-ffi/crt/aws-c-common/include/aws/common/private/system_info_priv.h +37 -0
- data/aws-crt-ffi/crt/aws-c-common/include/aws/common/system_info.h +47 -0
- data/aws-crt-ffi/crt/aws-c-common/include/aws/common/system_resource_util.h +30 -0
- data/aws-crt-ffi/crt/aws-c-common/include/aws/testing/aws_test_harness.h +3 -2
- data/aws-crt-ffi/crt/aws-c-common/source/allocator.c +64 -13
- data/aws-crt-ffi/crt/aws-c-common/source/android/logging.c +14 -0
- data/aws-crt-ffi/crt/aws-c-common/source/common.c +3 -3
- data/aws-crt-ffi/crt/aws-c-common/source/file.c +96 -35
- data/aws-crt-ffi/crt/aws-c-common/source/linux/system_info.c +24 -0
- data/aws-crt-ffi/crt/aws-c-common/source/memtrace.c +10 -3
- data/aws-crt-ffi/crt/aws-c-common/source/platform_fallback_stubs/system_info.c +21 -0
- data/aws-crt-ffi/crt/aws-c-common/source/posix/cross_process_lock.c +141 -0
- data/aws-crt-ffi/crt/aws-c-common/source/posix/system_info.c +1 -1
- data/aws-crt-ffi/crt/aws-c-common/source/posix/system_resource_utils.c +32 -0
- data/aws-crt-ffi/crt/aws-c-common/source/priority_queue.c +24 -0
- data/aws-crt-ffi/crt/aws-c-common/source/system_info.c +80 -0
- data/aws-crt-ffi/crt/aws-c-common/source/task_scheduler.c +2 -2
- data/aws-crt-ffi/crt/aws-c-common/source/windows/cross_process_lock.c +93 -0
- data/aws-crt-ffi/crt/aws-c-common/source/windows/system_resource_utils.c +31 -0
- data/aws-crt-ffi/crt/aws-c-common/tests/CMakeLists.txt +16 -0
- data/aws-crt-ffi/crt/aws-c-common/tests/alloc_test.c +83 -22
- data/aws-crt-ffi/crt/aws-c-common/tests/cross_process_lock_tests.c +116 -0
- data/aws-crt-ffi/crt/aws-c-common/tests/file_test.c +103 -0
- data/aws-crt-ffi/crt/aws-c-common/tests/priority_queue_test.c +36 -0
- data/aws-crt-ffi/crt/aws-c-common/tests/system_info_tests.c +19 -0
- data/aws-crt-ffi/crt/aws-c-common/tests/system_resource_util_test.c +37 -0
- data/aws-crt-ffi/crt/aws-c-http/include/aws/http/connection.h +9 -0
- data/aws-crt-ffi/crt/aws-c-http/include/aws/http/http.h +1 -0
- data/aws-crt-ffi/crt/aws-c-http/include/aws/http/private/connection_impl.h +5 -4
- data/aws-crt-ffi/crt/aws-c-http/include/aws/http/private/connection_manager_system_vtable.h +10 -18
- data/aws-crt-ffi/crt/aws-c-http/include/aws/http/private/proxy_impl.h +5 -1
- data/aws-crt-ffi/crt/aws-c-http/include/aws/http/private/request_response_impl.h +5 -0
- data/aws-crt-ffi/crt/aws-c-http/include/aws/http/request_response.h +10 -0
- data/aws-crt-ffi/crt/aws-c-http/source/connection.c +5 -2
- data/aws-crt-ffi/crt/aws-c-http/source/connection_manager.c +22 -21
- data/aws-crt-ffi/crt/aws-c-http/source/h1_connection.c +102 -17
- data/aws-crt-ffi/crt/aws-c-http/source/h1_stream.c +1 -0
- data/aws-crt-ffi/crt/aws-c-http/source/http.c +3 -0
- data/aws-crt-ffi/crt/aws-c-http/source/proxy_connection.c +2 -2
- data/aws-crt-ffi/crt/aws-c-http/tests/CMakeLists.txt +2 -0
- data/aws-crt-ffi/crt/aws-c-http/tests/test_connection_manager.c +18 -18
- data/aws-crt-ffi/crt/aws-c-http/tests/test_h1_client.c +111 -1
- data/aws-crt-ffi/crt/aws-c-http/tests/test_proxy.c +2 -2
- data/aws-crt-ffi/crt/aws-c-http/tests/test_stream_manager.c +2 -2
- data/aws-crt-ffi/crt/aws-c-io/include/aws/io/retry_strategy.h +1 -1
- data/aws-crt-ffi/crt/aws-c-io/source/exponential_backoff_retry_strategy.c +1 -1
- data/aws-crt-ffi/crt/aws-c-io/source/pkcs11_tls_op_handler.c +2 -4
- data/aws-crt-ffi/crt/aws-lc/CMakeLists.txt +16 -8
- data/aws-crt-ffi/crt/aws-lc/cmake/go.cmake +6 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/CMakeLists.txt +6 -9
- data/aws-crt-ffi/crt/aws-lc/crypto/asn1/a_time.c +34 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/asn1/a_utctm.c +4 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/asn1/asn1_test.cc +41 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/bio/bio_mem.c +6 -7
- data/aws-crt-ffi/crt/aws-lc/crypto/bio/bio_test.cc +152 -16
- data/aws-crt-ffi/crt/aws-lc/crypto/bio/connect.c +6 -12
- data/aws-crt-ffi/crt/aws-lc/crypto/bio/fd.c +2 -2
- data/aws-crt-ffi/crt/aws-lc/crypto/bio/file.c +20 -8
- data/aws-crt-ffi/crt/aws-lc/crypto/bio/socket.c +2 -2
- data/aws-crt-ffi/crt/aws-lc/crypto/bio/socket_helper.c +2 -2
- data/aws-crt-ffi/crt/aws-lc/crypto/blake2/blake2.c +11 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/bytestring/cbb.c +13 -3
- data/aws-crt-ffi/crt/aws-lc/crypto/bytestring/cbs.c +9 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/chacha/asm/chacha-armv8.pl +1 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/chacha/chacha.c +49 -8
- data/aws-crt-ffi/crt/aws-lc/crypto/chacha/chacha_test.cc +110 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/chacha/internal.h +8 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/compiler_test.cc +4 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/conf/conf_test.cc +1 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/crypto_test.cc +9 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/curve25519/curve25519.c +189 -108
- data/aws-crt-ffi/crt/aws-lc/crypto/curve25519/curve25519_nohw.c +78 -6
- data/aws-crt-ffi/crt/aws-lc/crypto/curve25519/ed25519_test.cc +9 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/curve25519/internal.h +24 -10
- data/aws-crt-ffi/crt/aws-lc/crypto/curve25519/spake25519.c +4 -4
- data/aws-crt-ffi/crt/aws-lc/crypto/curve25519/x25519_test.cc +80 -11
- data/aws-crt-ffi/crt/aws-lc/crypto/decrepit/evp/evp_do_all.c +2 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/digest_extra/digest_extra.c +8 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/digest_extra/digest_test.cc +110 -45
- data/aws-crt-ffi/crt/aws-lc/crypto/dsa/dsa_test.cc +8 -2
- data/aws-crt-ffi/crt/aws-lc/crypto/dsa/internal.h +18 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/dynamic_loading_test.c +8 -5
- data/aws-crt-ffi/crt/aws-lc/crypto/ec_extra/ec_derive.c +4 -3
- data/aws-crt-ffi/crt/aws-lc/crypto/ec_extra/hash_to_curve.c +6 -18
- data/aws-crt-ffi/crt/aws-lc/crypto/endian_test.cc +308 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/err/ssl.errordata +2 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/evp_extra/evp_extra_test.cc +2 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/evp_extra/evp_test.cc +11 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/evp_extra/evp_tests.txt +25 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/evp_extra/p_ec_asn1.c +1 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/evp_extra/p_kem.c +2 -2
- data/aws-crt-ffi/crt/aws-lc/crypto/evp_extra/p_rsa_asn1.c +1 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/evp_extra/print.c +7 -6
- data/aws-crt-ffi/crt/aws-lc/crypto/evp_extra/scrypt.c +13 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/CMakeLists.txt +13 -4
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/aes/aes_nohw.c +18 -6
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/bcm.c +12 -4
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/bn/bn_assert_test.cc +77 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/bn/bn_test.cc +30 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/bn/bytes.c +112 -22
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/bn/div.c +12 -5
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/bn/exponentiation.c +54 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/bn/gcd.c +5 -6
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/bn/internal.h +37 -15
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/bn/montgomery.c +4 -11
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/bn/montgomery_inv.c +51 -15
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/cipher/aead.c +2 -2
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/digest/digest.c +29 -6
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/digest/digests.c +89 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/digest/internal.h +4 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/ec.c +19 -36
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/ec_key.c +3 -3
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/ec_montgomery.c +9 -7
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/ec_test.cc +33 -9
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/internal.h +17 -12
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/p224-64.c +5 -8
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/p256-nistz.c +8 -8
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/p256.c +9 -8
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/p384.c +33 -16
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/p521.c +14 -6
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/scalar.c +26 -24
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/simple_mul.c +8 -5
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ec/wnaf.c +3 -3
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/ecdsa/ecdsa.c +9 -3
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/evp/evp.c +43 -12
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/evp/p_ec.c +4 -3
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/hmac/hmac.c +3 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/modes/xts.c +26 -3
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/rand/cpu_jitter_test.cc +1 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/rand/internal.h +20 -11
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/rand/rand.c +10 -10
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/rand/urandom.c +2 -2
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/rsa/internal.h +59 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/rsa/padding.c +9 -3
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/rsa/rsa.c +7 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/rsa/rsa_impl.c +51 -60
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/service_indicator/service_indicator.c +5 -2
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/service_indicator/service_indicator_test.cc +205 -5
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/sha/asm/sha1-armv8.pl +1 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/sha/asm/sha512-armv8.pl +1 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/sha/internal.h +8 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/sha/sha3.c +37 -15
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/sha/sha3_test.cc +115 -110
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/sha/sha512.c +55 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/fipsmodule/sshkdf/sshkdf.c +2 -2
- data/aws-crt-ffi/crt/aws-lc/crypto/hmac_extra/hmac_test.cc +12 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/hmac_extra/hmac_tests.txt +10 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/hrss/asm/poly_rq_mul.S +2 -6
- data/aws-crt-ffi/crt/aws-lc/crypto/impl_dispatch_test.cc +9 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/internal.h +90 -8
- data/aws-crt-ffi/crt/aws-lc/crypto/kem/kem.c +28 -27
- data/aws-crt-ffi/crt/aws-lc/crypto/kyber/kem_kyber.h +14 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/obj/obj_dat.h +52 -2
- data/aws-crt-ffi/crt/aws-lc/crypto/obj/obj_mac.num +5 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/obj/objects.txt +7 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/perlasm/arm-xlate.pl +3 -14
- data/aws-crt-ffi/crt/aws-lc/crypto/perlasm/ppc-xlate.pl +1 -5
- data/aws-crt-ffi/crt/aws-lc/crypto/perlasm/x86_64-xlate.pl +4 -15
- data/aws-crt-ffi/crt/aws-lc/crypto/perlasm/x86asm.pl +4 -13
- data/aws-crt-ffi/crt/aws-lc/crypto/poly1305/poly1305_arm_asm.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/crypto/rand_extra/deterministic.c +4 -3
- data/aws-crt-ffi/crt/aws-lc/crypto/rand_extra/fuchsia.c +4 -4
- data/aws-crt-ffi/crt/aws-lc/crypto/rand_extra/rand_test.cc +0 -63
- data/aws-crt-ffi/crt/aws-lc/crypto/rand_extra/windows.c +41 -19
- data/aws-crt-ffi/crt/aws-lc/crypto/rsa_extra/rsa_test.cc +3 -3
- data/aws-crt-ffi/crt/aws-lc/crypto/siphash/siphash.c +12 -5
- data/aws-crt-ffi/crt/aws-lc/crypto/siphash/siphash_test.cc +5 -5
- data/aws-crt-ffi/crt/aws-lc/crypto/stack/stack.c +68 -46
- data/aws-crt-ffi/crt/aws-lc/crypto/trust_token/pmbtoken.c +4 -4
- data/aws-crt-ffi/crt/aws-lc/crypto/trust_token/voprf.c +2 -2
- data/aws-crt-ffi/crt/aws-lc/crypto/x509/by_dir.c +0 -6
- data/aws-crt-ffi/crt/aws-lc/crypto/x509/internal.h +4 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/x509/x509_lu.c +33 -9
- data/aws-crt-ffi/crt/aws-lc/crypto/x509/x509_test.cc +87 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/x509/x509_trs.c +1 -1
- data/aws-crt-ffi/crt/aws-lc/crypto/x509/x509_vfy.c +35 -13
- data/aws-crt-ffi/crt/aws-lc/crypto/x509v3/v3_lib.c +2 -0
- data/aws-crt-ffi/crt/aws-lc/crypto/x509v3/v3_purp.c +4 -6
- data/aws-crt-ffi/crt/aws-lc/generated-src/crypto_test_data.cc +179 -151
- data/aws-crt-ffi/crt/aws-lc/generated-src/err_data.c +353 -349
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/chacha/chacha-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/cipher_extra/chacha20_poly1305_armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/aesv8-armx.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/aesv8-gcm-armv8-unroll8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/aesv8-gcm-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/armv8-mont.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/bn-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/ghash-neon-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/ghashv8-armx.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/keccak1600-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/md5-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/p256-armv8-asm.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/p256_beeu-armv8-asm.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/sha1-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/sha256-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/sha512-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/fipsmodule/vpaes-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-aarch64/crypto/test/trampoline-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-arm/crypto/chacha/chacha-armv4.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-arm/crypto/fipsmodule/aesv8-armx.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-arm/crypto/fipsmodule/armv4-mont.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-arm/crypto/fipsmodule/bsaes-armv7.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-arm/crypto/fipsmodule/ghash-armv4.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-arm/crypto/fipsmodule/ghashv8-armx.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-arm/crypto/fipsmodule/sha1-armv4-large.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-arm/crypto/fipsmodule/sha256-armv4.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-arm/crypto/fipsmodule/sha512-armv4.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-arm/crypto/fipsmodule/vpaes-armv7.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/ios-arm/crypto/test/trampoline-armv4.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/chacha/chacha-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/cipher_extra/chacha20_poly1305_armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/aesv8-armx.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/aesv8-gcm-armv8-unroll8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/aesv8-gcm-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/armv8-mont.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/bn-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/ghash-neon-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/ghashv8-armx.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/keccak1600-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/md5-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/p256-armv8-asm.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/p256_beeu-armv8-asm.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/sha1-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/sha256-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/sha512-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/fipsmodule/vpaes-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-aarch64/crypto/test/trampoline-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-arm/crypto/chacha/chacha-armv4.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-arm/crypto/fipsmodule/aesv8-armx.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-arm/crypto/fipsmodule/armv4-mont.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-arm/crypto/fipsmodule/bsaes-armv7.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-arm/crypto/fipsmodule/ghash-armv4.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-arm/crypto/fipsmodule/ghashv8-armx.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-arm/crypto/fipsmodule/sha1-armv4-large.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-arm/crypto/fipsmodule/sha256-armv4.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-arm/crypto/fipsmodule/sha512-armv4.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-arm/crypto/fipsmodule/vpaes-armv7.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-arm/crypto/test/trampoline-armv4.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-ppc64le/crypto/fipsmodule/aesp8-ppc.S +1 -5
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-ppc64le/crypto/fipsmodule/ghashp8-ppc.S +1 -5
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-ppc64le/crypto/test/trampoline-ppc.S +1 -5
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/chacha/chacha-x86.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/fipsmodule/aesni-x86.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/fipsmodule/bn-586.S +4 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/fipsmodule/co-586.S +4 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/fipsmodule/ghash-ssse3-x86.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/fipsmodule/ghash-x86.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/fipsmodule/md5-586.S +4 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/fipsmodule/sha1-586.S +4 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/fipsmodule/sha256-586.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/fipsmodule/sha512-586.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/fipsmodule/vpaes-x86.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/fipsmodule/x86-mont.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86/crypto/test/trampoline-x86.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/chacha/chacha-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/cipher_extra/aes128gcmsiv-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/cipher_extra/aesni-sha1-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/cipher_extra/aesni-sha256-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/cipher_extra/chacha20_poly1305_x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/aesni-gcm-avx512.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/aesni-gcm-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/aesni-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/aesni-xts-avx512.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/ghash-ssse3-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/ghash-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/md5-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/p256-x86_64-asm.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/p256_beeu-x86_64-asm.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/rdrand-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/rsaz-avx2.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/sha1-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/sha256-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/sha512-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/vpaes-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/x86_64-mont.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/fipsmodule/x86_64-mont5.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/linux-x86_64/crypto/test/trampoline-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/chacha/chacha-x86.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/fipsmodule/aesni-x86.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/fipsmodule/bn-586.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/fipsmodule/co-586.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/fipsmodule/ghash-ssse3-x86.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/fipsmodule/ghash-x86.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/fipsmodule/md5-586.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/fipsmodule/sha1-586.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/fipsmodule/sha256-586.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/fipsmodule/sha512-586.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/fipsmodule/vpaes-x86.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/fipsmodule/x86-mont.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86/crypto/test/trampoline-x86.S +3 -12
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/chacha/chacha-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/cipher_extra/aes128gcmsiv-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/cipher_extra/aesni-sha1-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/cipher_extra/aesni-sha256-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/cipher_extra/chacha20_poly1305_x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/aesni-gcm-avx512.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/aesni-gcm-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/aesni-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/aesni-xts-avx512.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/ghash-ssse3-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/ghash-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/md5-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/p256-x86_64-asm.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/p256_beeu-x86_64-asm.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/rdrand-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/rsaz-avx2.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/sha1-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/sha256-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/sha512-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/vpaes-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/x86_64-mont.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/fipsmodule/x86_64-mont5.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/mac-x86_64/crypto/test/trampoline-x86_64.S +2 -11
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/chacha/chacha-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/cipher_extra/chacha20_poly1305_armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/aesv8-armx.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/aesv8-gcm-armv8-unroll8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/aesv8-gcm-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/armv8-mont.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/bn-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/ghash-neon-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/ghashv8-armx.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/keccak1600-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/md5-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/p256-armv8-asm.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/p256_beeu-armv8-asm.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/sha1-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/sha256-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/sha512-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/fipsmodule/vpaes-armv8.S +3 -13
- data/aws-crt-ffi/crt/aws-lc/generated-src/win-aarch64/crypto/test/trampoline-armv8.S +4 -14
- data/aws-crt-ffi/crt/aws-lc/go.mod +4 -4
- data/aws-crt-ffi/crt/aws-lc/go.sum +8 -10
- data/aws-crt-ffi/crt/aws-lc/include/openssl/aead.h +2 -2
- data/aws-crt-ffi/crt/aws-lc/include/openssl/arm_arch.h +4 -119
- data/aws-crt-ffi/crt/aws-lc/include/openssl/asm_base.h +185 -0
- data/aws-crt-ffi/crt/aws-lc/include/openssl/asn1.h +5 -0
- data/aws-crt-ffi/crt/aws-lc/include/openssl/base.h +31 -134
- data/aws-crt-ffi/crt/aws-lc/include/openssl/bio.h +30 -18
- data/aws-crt-ffi/crt/aws-lc/include/openssl/bn.h +0 -2
- data/aws-crt-ffi/crt/aws-lc/include/openssl/chacha.h +6 -0
- data/aws-crt-ffi/crt/aws-lc/include/openssl/cipher.h +2 -2
- data/aws-crt-ffi/crt/aws-lc/include/openssl/digest.h +9 -6
- data/aws-crt-ffi/crt/aws-lc/include/openssl/dsa.h +0 -21
- data/aws-crt-ffi/crt/aws-lc/include/openssl/ec.h +1 -1
- data/aws-crt-ffi/crt/aws-lc/include/openssl/err.h +1 -1
- data/aws-crt-ffi/crt/aws-lc/include/openssl/evp.h +8 -5
- data/aws-crt-ffi/crt/aws-lc/include/openssl/nid.h +21 -0
- data/aws-crt-ffi/crt/aws-lc/include/openssl/rsa.h +1 -65
- data/aws-crt-ffi/crt/aws-lc/include/openssl/sha.h +22 -1
- data/aws-crt-ffi/crt/aws-lc/include/openssl/ssl.h +121 -13
- data/aws-crt-ffi/crt/aws-lc/include/openssl/stack.h +229 -208
- data/aws-crt-ffi/crt/aws-lc/include/openssl/target.h +166 -0
- data/aws-crt-ffi/crt/aws-lc/include/openssl/x509.h +30 -10
- data/aws-crt-ffi/crt/aws-lc/include/openssl/x509v3.h +6 -4
- data/aws-crt-ffi/crt/aws-lc/sources.cmake +2 -0
- data/aws-crt-ffi/crt/aws-lc/ssl/extensions.cc +12 -7
- data/aws-crt-ffi/crt/aws-lc/ssl/handshake_server.cc +28 -18
- data/aws-crt-ffi/crt/aws-lc/ssl/internal.h +41 -6
- data/aws-crt-ffi/crt/aws-lc/ssl/s3_both.cc +9 -17
- data/aws-crt-ffi/crt/aws-lc/ssl/ssl_cipher.cc +13 -5
- data/aws-crt-ffi/crt/aws-lc/ssl/ssl_key_share.cc +542 -2
- data/aws-crt-ffi/crt/aws-lc/ssl/ssl_lib.cc +35 -0
- data/aws-crt-ffi/crt/aws-lc/ssl/ssl_test.cc +1847 -14
- data/aws-crt-ffi/crt/aws-lc/ssl/ssl_x509.cc +128 -0
- data/aws-crt-ffi/crt/aws-lc/ssl/test/PORTING.md +10 -7
- data/aws-crt-ffi/crt/aws-lc/ssl/test/bssl_shim.cc +133 -77
- data/aws-crt-ffi/crt/aws-lc/ssl/test/handshake_util.cc +3 -3
- data/aws-crt-ffi/crt/aws-lc/ssl/test/handshaker.cc +4 -0
- data/aws-crt-ffi/crt/aws-lc/ssl/test/runner/handshake_client.go +6 -2
- data/aws-crt-ffi/crt/aws-lc/ssl/test/runner/handshake_messages.go +894 -1042
- data/aws-crt-ffi/crt/aws-lc/ssl/test/runner/handshake_server.go +24 -23
- data/aws-crt-ffi/crt/aws-lc/ssl/test/runner/prf.go +6 -5
- data/aws-crt-ffi/crt/aws-lc/ssl/test/runner/runner.go +56 -55
- data/aws-crt-ffi/crt/aws-lc/ssl/test/runner/shim_dispatcher.go +188 -0
- data/aws-crt-ffi/crt/aws-lc/ssl/test/runner/ticket.go +37 -39
- data/aws-crt-ffi/crt/aws-lc/ssl/test/test_config.cc +59 -24
- data/aws-crt-ffi/crt/aws-lc/ssl/test/test_config.h +3 -2
- data/aws-crt-ffi/crt/aws-lc/ssl/tls13_server.cc +10 -11
- data/aws-crt-ffi/crt/aws-lc/tests/ci/cdk/app.py +4 -4
- data/aws-crt-ffi/crt/aws-lc/tests/ci/cdk/cdk/{aws_lc_mac_arm_ci_stack.py → aws_lc_ec2_test_framework_ci_stack.py} +13 -29
- data/aws-crt-ffi/crt/aws-lc/tests/ci/cdk/cdk/ssm/general_test_run_ssm_document.yaml +43 -0
- data/aws-crt-ffi/crt/aws-lc/tests/ci/common_posix_setup.sh +10 -0
- data/aws-crt-ffi/crt/aws-lc/tests/ci/docker_images/linux-aarch/amazonlinux-2023_base/Dockerfile +5 -1
- data/aws-crt-ffi/crt/aws-lc/tests/ci/docker_images/linux-aarch/ubuntu-22.04_base/Dockerfile +19 -3
- data/aws-crt-ffi/crt/aws-lc/tests/ci/docker_images/linux-x86/amazonlinux-2_gcc-7x-intel-sde/Dockerfile +5 -4
- data/aws-crt-ffi/crt/aws-lc/tests/ci/docker_images/linux-x86/build_images.sh +1 -0
- data/aws-crt-ffi/crt/aws-lc/tests/ci/docker_images/linux-x86/push_images.sh +2 -1
- data/aws-crt-ffi/crt/aws-lc/tests/ci/docker_images/linux-x86/ubuntu-20.04_clang-10x_formal-verification/create_image.sh +1 -1
- data/aws-crt-ffi/crt/aws-lc/tests/ci/docker_images/linux-x86/ubuntu-22.04_base/Dockerfile +1 -0
- data/aws-crt-ffi/crt/aws-lc/tests/ci/docker_images/linux-x86/ubuntu-22.04_clang-14x-sde/Dockerfile +42 -0
- data/aws-crt-ffi/crt/aws-lc/tests/ci/docker_images/windows/vs2017/Dockerfile +14 -0
- data/aws-crt-ffi/crt/aws-lc/tests/ci/docker_images/windows/windows_base/Dockerfile +3 -0
- data/aws-crt-ffi/crt/aws-lc/tests/ci/integration/README.md +12 -0
- data/aws-crt-ffi/crt/aws-lc/tests/ci/integration/nginx_patch/aws-lc-nginx.patch +68 -23
- data/aws-crt-ffi/crt/aws-lc/tests/ci/integration/run_crt_integration.sh +27 -0
- data/aws-crt-ffi/crt/aws-lc/tests/ci/integration/run_monit_integration.sh +56 -0
- data/aws-crt-ffi/crt/aws-lc/tests/ci/integration/sslproxy_patch/aws-lc-sslproxy.patch +2 -2
- data/aws-crt-ffi/crt/aws-lc/tests/ci/run_ec2_test_framework.sh +135 -0
- data/aws-crt-ffi/crt/aws-lc/tests/ci/run_fips_tests.sh +14 -2
- data/aws-crt-ffi/crt/aws-lc/tests/ci/run_tests_with_sde.sh +4 -1
- data/aws-crt-ffi/crt/aws-lc/tests/ci/run_tests_with_sde_asan.sh +14 -0
- data/aws-crt-ffi/crt/aws-lc/tests/ci/run_windows_tests.bat +39 -3
- data/aws-crt-ffi/crt/aws-lc/third_party/fiat/README.md +21 -6
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/bignum_madd_n25519.S +284 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/bignum_madd_n25519_alt.S +210 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/bignum_mod_n25519.S +186 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/bignum_neg_p25519.S +65 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/curve25519_x25519.S +1043 -352
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/curve25519_x25519_alt.S +1043 -352
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/curve25519_x25519_byte.S +1043 -352
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/curve25519_x25519_byte_alt.S +1043 -352
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/curve25519_x25519base.S +1042 -352
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/curve25519_x25519base_alt.S +1042 -352
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/curve25519_x25519base_byte.S +1042 -352
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/curve25519_x25519base_byte_alt.S +1043 -354
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/edwards25519_decode.S +700 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/edwards25519_decode_alt.S +563 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/edwards25519_encode.S +131 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/edwards25519_scalarmulbase.S +9626 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/edwards25519_scalarmulbase_alt.S +9468 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/edwards25519_scalarmuldouble.S +3157 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/curve25519/edwards25519_scalarmuldouble_alt.S +2941 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/p384/Makefile +1 -1
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/arm/p521/Makefile +1 -1
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/include/s2n-bignum_aws-lc.h +34 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/bignum_madd_n25519.S +219 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/bignum_madd_n25519_alt.S +245 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/bignum_mod_n25519.S +228 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/bignum_neg_p25519.S +86 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/curve25519_x25519.S +1350 -407
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/curve25519_x25519_alt.S +1350 -407
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/curve25519_x25519base.S +1344 -400
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/curve25519_x25519base_alt.S +1348 -402
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/edwards25519_decode.S +670 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/edwards25519_decode_alt.S +751 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/edwards25519_encode.S +81 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/edwards25519_scalarmulbase.S +9910 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/edwards25519_scalarmulbase_alt.S +9986 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/edwards25519_scalarmuldouble.S +3619 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/edwards25519_scalarmuldouble_alt.S +3736 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/wycheproof_testvectors/hmac_sha512_224_test.json +1978 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/wycheproof_testvectors/hmac_sha512_224_test.txt +1403 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/wycheproof_testvectors/hmac_sha512_256_test.json +1993 -0
- data/aws-crt-ffi/crt/aws-lc/third_party/wycheproof_testvectors/hmac_sha512_256_test.txt +1416 -0
- data/aws-crt-ffi/crt/aws-lc/tool/digest.cc +4 -0
- data/aws-crt-ffi/crt/aws-lc/tool/internal.h +1 -0
- data/aws-crt-ffi/crt/aws-lc/tool/speed.cc +53 -6
- data/aws-crt-ffi/crt/aws-lc/util/all_tests.go +43 -12
- data/aws-crt-ffi/crt/aws-lc/util/all_tests.json +13 -5
- data/aws-crt-ffi/crt/aws-lc/util/bot/DEPS +4 -4
- data/aws-crt-ffi/crt/aws-lc/util/bot/update_clang.py +8 -2
- data/aws-crt-ffi/crt/aws-lc/util/codecov-ci.sh +82 -0
- data/aws-crt-ffi/crt/aws-lc/util/convert_wycheproof/convert_wycheproof.go +7 -5
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/ACVP.md +7 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/subprocess/hash.go +24 -9
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/subprocess/rsa.go +3 -4
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/subprocess/subprocess.go +15 -10
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/test/expected/HMAC-SHA2-512-224.bz2 +0 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/test/expected/SHA2-512-224.bz2 +0 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/test/expected/SHAKE-128.bz2 +0 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/test/expected/SHAKE-256.bz2 +0 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/test/sha-tests/sha512-224-tests.json +1 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/test/sha-tests/shake-128-tests.json +1 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/test/sha-tests/shake-256-tests.json +1 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/test/tests.json +1 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/test/vectors/HMAC-SHA2-512-224.bz2 +0 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/test/vectors/SHA2-512-224.bz2 +0 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/test/vectors/SHAKE-128.bz2 +0 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/acvptool/test/vectors/SHAKE-256.bz2 +0 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/modulewrapper/main.cc +4 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/acvp/modulewrapper/modulewrapper.cc +144 -1
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/delocate/delocate.go +9 -3
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/delocate/testdata/aarch64-Basic/in.s +4 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/delocate/testdata/aarch64-Basic/out.s +11 -0
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/inject_hash/inject_hash.go +13 -4
- data/aws-crt-ffi/crt/aws-lc/util/fipstools/test-break-kat.sh +2 -0
- data/aws-crt-ffi/crt/aws-lc/util/testconfig/testconfig.go +2 -1
- data/aws-crt-ffi/crt/s2n/api/s2n.h +9 -5
- data/aws-crt-ffi/crt/s2n/bindings/rust/bench/benches/handshake.rs +9 -6
- data/aws-crt-ffi/crt/s2n/bindings/rust/bench/benches/resumption.rs +14 -14
- data/aws-crt-ffi/crt/s2n/bindings/rust/bench/benches/throughput.rs +9 -6
- data/aws-crt-ffi/crt/s2n/bindings/rust/bench/src/harness.rs +106 -102
- data/aws-crt-ffi/crt/s2n/bindings/rust/bench/src/openssl.rs +24 -20
- data/aws-crt-ffi/crt/s2n/bindings/rust/bench/src/rustls.rs +28 -24
- data/aws-crt-ffi/crt/s2n/bindings/rust/bench/src/s2n_tls.rs +52 -50
- data/aws-crt-ffi/crt/s2n/bindings/rust/generate/Cargo.toml +1 -0
- data/aws-crt-ffi/crt/s2n/bindings/rust/integration/Cargo.toml +3 -0
- data/aws-crt-ffi/crt/s2n/bindings/rust/s2n-tls/Cargo.toml +2 -2
- data/aws-crt-ffi/crt/s2n/bindings/rust/s2n-tls/src/connection.rs +9 -0
- data/aws-crt-ffi/crt/s2n/bindings/rust/s2n-tls-sys/templates/Cargo.template +2 -1
- data/aws-crt-ffi/crt/s2n/bindings/rust/s2n-tls-tokio/Cargo.toml +2 -2
- data/aws-crt-ffi/crt/s2n/tests/cbmc/sources/make_common_datastructures.c +9 -2
- data/aws-crt-ffi/crt/s2n/tests/fuzz/s2n_client_cert_verify_recv_test.c +1 -1
- data/aws-crt-ffi/crt/s2n/tests/fuzz/s2n_hybrid_ecdhe_kyber_r3_fuzz_test.c +1 -1
- data/aws-crt-ffi/crt/s2n/tests/fuzz/s2n_tls13_cert_verify_recv_test.c +1 -1
- data/aws-crt-ffi/crt/s2n/tests/integrationv2/test_version_negotiation.py +4 -4
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_auth_selection_test.c +19 -9
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_client_auth_handshake_test.c +3 -3
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_client_cert_verify_test.c +1 -1
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_client_hello_recv_test.c +1 -1
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_client_hello_test.c +4 -4
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_client_signature_algorithms_extension_test.c +4 -5
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_connection_protocol_versions_test.c +390 -0
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_connection_test.c +8 -4
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_handshake_test.c +2 -1
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_quic_support_io_test.c +106 -0
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_security_policies_test.c +6 -2
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_self_talk_offload_signing_test.c +3 -3
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_self_talk_session_resumption_test.c +135 -0
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_server_new_session_ticket_test.c +32 -0
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_server_signature_algorithms_extension_test.c +1 -1
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_signature_algorithms_test.c +307 -283
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_tls13_cert_request_test.c +1 -1
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_tls13_cert_verify_test.c +18 -17
- data/aws-crt-ffi/crt/s2n/tests/unit/s2n_x509_validator_test.c +125 -0
- data/aws-crt-ffi/crt/s2n/tls/extensions/s2n_client_signature_algorithms.c +8 -1
- data/aws-crt-ffi/crt/s2n/tls/extensions/s2n_client_supported_versions.c +43 -11
- data/aws-crt-ffi/crt/s2n/tls/extensions/s2n_client_supported_versions.h +3 -0
- data/aws-crt-ffi/crt/s2n/tls/extensions/s2n_server_signature_algorithms.c +8 -1
- data/aws-crt-ffi/crt/s2n/tls/s2n_auth_selection.c +4 -2
- data/aws-crt-ffi/crt/s2n/tls/s2n_client_cert_verify.c +7 -10
- data/aws-crt-ffi/crt/s2n/tls/s2n_client_hello.c +2 -2
- data/aws-crt-ffi/crt/s2n/tls/s2n_connection.c +75 -14
- data/aws-crt-ffi/crt/s2n/tls/s2n_handshake.h +2 -2
- data/aws-crt-ffi/crt/s2n/tls/s2n_post_handshake.c +1 -1
- data/aws-crt-ffi/crt/s2n/tls/s2n_post_handshake.h +1 -0
- data/aws-crt-ffi/crt/s2n/tls/s2n_quic_support.c +29 -0
- data/aws-crt-ffi/crt/s2n/tls/s2n_quic_support.h +5 -0
- data/aws-crt-ffi/crt/s2n/tls/s2n_security_policies.c +40 -0
- data/aws-crt-ffi/crt/s2n/tls/s2n_security_policies.h +4 -0
- data/aws-crt-ffi/crt/s2n/tls/s2n_server_cert_request.c +1 -1
- data/aws-crt-ffi/crt/s2n/tls/s2n_server_hello.c +0 -3
- data/aws-crt-ffi/crt/s2n/tls/s2n_server_key_exchange.c +8 -9
- data/aws-crt-ffi/crt/s2n/tls/s2n_server_new_session_ticket.c +8 -0
- data/aws-crt-ffi/crt/s2n/tls/s2n_signature_algorithms.c +111 -72
- data/aws-crt-ffi/crt/s2n/tls/s2n_signature_algorithms.h +11 -9
- data/aws-crt-ffi/crt/s2n/tls/s2n_signature_scheme.c +9 -0
- data/aws-crt-ffi/crt/s2n/tls/s2n_signature_scheme.h +2 -0
- data/aws-crt-ffi/crt/s2n/tls/s2n_tls13_certificate_verify.c +12 -18
- data/aws-crt-ffi/crt/s2n/tls/s2n_x509_validator.c +7 -7
- data/aws-crt-ffi/src/api.h +1 -0
- data/lib/aws-crt/native.rb +1 -1
- metadata +68 -5
- data/aws-crt-ffi/crt/aws-lc/tests/ci/cdk/cdk/ssm/m1_tests_ssm_document.yaml +0 -34
- data/aws-crt-ffi/crt/aws-lc/tests/ci/run_m1_ec2_instance.sh +0 -96
data/aws-crt-ffi/crt/aws-lc/third_party/s2n-bignum/x86_att/curve25519/edwards25519_scalarmuldouble.S
ADDED
|
@@ -0,0 +1,3619 @@
|
|
|
1
|
+
// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
|
|
2
|
+
// SPDX-License-Identifier: Apache-2.0 OR ISC
|
|
3
|
+
|
|
4
|
+
// ----------------------------------------------------------------------------
|
|
5
|
+
// Double scalar multiplication for edwards25519, fresh and base point
|
|
6
|
+
// Input scalar[4], point[8], bscalar[4]; output res[8]
|
|
7
|
+
//
|
|
8
|
+
// extern void edwards25519_scalarmuldouble
|
|
9
|
+
// (uint64_t res[static 8],uint64_t scalar[static 4],
|
|
10
|
+
// uint64_t point[static 8],uint64_t bscalar[static 4]);
|
|
11
|
+
//
|
|
12
|
+
// Given scalar = n, point = P and bscalar = m, returns in res
|
|
13
|
+
// the point (X,Y) = n * P + m * B where B = (...,4/5) is
|
|
14
|
+
// the standard basepoint for the edwards25519 (Ed25519) curve.
|
|
15
|
+
//
|
|
16
|
+
// Both 256-bit coordinates of the input point P are implicitly
|
|
17
|
+
// reduced modulo 2^255-19 if they are not already in reduced form,
|
|
18
|
+
// but the conventional usage is that they *are* already reduced.
|
|
19
|
+
// The scalars can be arbitrary 256-bit numbers but may also be
|
|
20
|
+
// considered as implicitly reduced modulo the group order.
|
|
21
|
+
//
|
|
22
|
+
// Standard x86-64 ABI: RDI = res, RSI = scalar, RDX = point, RCX = bscalar
|
|
23
|
+
// Microsoft x64 ABI: RCX = res, RDX = scalar, R8 = point, R9 = bscalar
|
|
24
|
+
// ----------------------------------------------------------------------------
|
|
25
|
+
#include "_internal_s2n_bignum.h"
|
|
26
|
+
|
|
27
|
+
|
|
28
|
+
S2N_BN_SYM_VISIBILITY_DIRECTIVE(edwards25519_scalarmuldouble)
|
|
29
|
+
S2N_BN_SYM_PRIVACY_DIRECTIVE(edwards25519_scalarmuldouble)
|
|
30
|
+
.text
|
|
31
|
+
|
|
32
|
+
// Size of individual field elements
|
|
33
|
+
|
|
34
|
+
#define NUMSIZE 32
|
|
35
|
+
|
|
36
|
+
// Pointer-offset pairs for result and temporaries on stack with some aliasing.
|
|
37
|
+
// Both "resx" and "resy" assume the "res" pointer has been preloaded into %rbp.
|
|
38
|
+
|
|
39
|
+
#define resx (0*NUMSIZE)(%rbp)
|
|
40
|
+
#define resy (1*NUMSIZE)(%rbp)
|
|
41
|
+
|
|
42
|
+
#define scalar (0*NUMSIZE)(%rsp)
|
|
43
|
+
#define bscalar (1*NUMSIZE)(%rsp)
|
|
44
|
+
|
|
45
|
+
#define tabent (2*NUMSIZE)(%rsp)
|
|
46
|
+
#define btabent (6*NUMSIZE)(%rsp)
|
|
47
|
+
|
|
48
|
+
#define acc (9*NUMSIZE)(%rsp)
|
|
49
|
+
|
|
50
|
+
#define tab (13*NUMSIZE)(%rsp)
|
|
51
|
+
|
|
52
|
+
// Additional variables kept on the stack
|
|
53
|
+
|
|
54
|
+
#define bf 45*NUMSIZE(%rsp)
|
|
55
|
+
#define cf 45*NUMSIZE+8(%rsp)
|
|
56
|
+
#define i 45*NUMSIZE+16(%rsp)
|
|
57
|
+
#define res 45*NUMSIZE+24(%rsp)
|
|
58
|
+
|
|
59
|
+
// Total size to reserve on the stack (excluding local subroutines)
|
|
60
|
+
|
|
61
|
+
#define NSPACE (46*NUMSIZE)
|
|
62
|
+
|
|
63
|
+
// Syntactic variants to make x86_att forms easier to generate
|
|
64
|
+
|
|
65
|
+
#define SCALAR (0*NUMSIZE)
|
|
66
|
+
#define BSCALAR (1*NUMSIZE)
|
|
67
|
+
#define TABENT (2*NUMSIZE)
|
|
68
|
+
#define BTABENT (6*NUMSIZE)
|
|
69
|
+
#define ACC (9*NUMSIZE)
|
|
70
|
+
#define TAB (13*NUMSIZE)
|
|
71
|
+
|
|
72
|
+
// Sub-references used in local subroutines with local stack
|
|
73
|
+
|
|
74
|
+
#define x_0 0(%rdi)
|
|
75
|
+
#define y_0 NUMSIZE(%rdi)
|
|
76
|
+
#define z_0 (2*NUMSIZE)(%rdi)
|
|
77
|
+
#define w_0 (3*NUMSIZE)(%rdi)
|
|
78
|
+
|
|
79
|
+
#define x_1 0(%rsi)
|
|
80
|
+
#define y_1 NUMSIZE(%rsi)
|
|
81
|
+
#define z_1 (2*NUMSIZE)(%rsi)
|
|
82
|
+
#define w_1 (3*NUMSIZE)(%rsi)
|
|
83
|
+
|
|
84
|
+
#define x_2 0(%rbp)
|
|
85
|
+
#define y_2 NUMSIZE(%rbp)
|
|
86
|
+
#define z_2 (2*NUMSIZE)(%rbp)
|
|
87
|
+
#define w_2 (3*NUMSIZE)(%rbp)
|
|
88
|
+
|
|
89
|
+
#define t0 (0*NUMSIZE)(%rsp)
|
|
90
|
+
#define t1 (1*NUMSIZE)(%rsp)
|
|
91
|
+
#define t2 (2*NUMSIZE)(%rsp)
|
|
92
|
+
#define t3 (3*NUMSIZE)(%rsp)
|
|
93
|
+
#define t4 (4*NUMSIZE)(%rsp)
|
|
94
|
+
#define t5 (5*NUMSIZE)(%rsp)
|
|
95
|
+
|
|
96
|
+
// Macro wrapping up the basic field multiplication, only trivially
|
|
97
|
+
// different from a pure function call to bignum_mul_p25519.
|
|
98
|
+
|
|
99
|
+
#define mul_p25519(P0,P1,P2) \
|
|
100
|
+
xorl %ecx, %ecx ; \
|
|
101
|
+
movq P2, %rdx ; \
|
|
102
|
+
mulxq P1, %r8, %r9 ; \
|
|
103
|
+
mulxq 0x8+P1, %rax, %r10 ; \
|
|
104
|
+
addq %rax, %r9 ; \
|
|
105
|
+
mulxq 0x10+P1, %rax, %r11 ; \
|
|
106
|
+
adcq %rax, %r10 ; \
|
|
107
|
+
mulxq 0x18+P1, %rax, %r12 ; \
|
|
108
|
+
adcq %rax, %r11 ; \
|
|
109
|
+
adcq %rcx, %r12 ; \
|
|
110
|
+
xorl %ecx, %ecx ; \
|
|
111
|
+
movq 0x8+P2, %rdx ; \
|
|
112
|
+
mulxq P1, %rax, %rbx ; \
|
|
113
|
+
adcxq %rax, %r9 ; \
|
|
114
|
+
adoxq %rbx, %r10 ; \
|
|
115
|
+
mulxq 0x8+P1, %rax, %rbx ; \
|
|
116
|
+
adcxq %rax, %r10 ; \
|
|
117
|
+
adoxq %rbx, %r11 ; \
|
|
118
|
+
mulxq 0x10+P1, %rax, %rbx ; \
|
|
119
|
+
adcxq %rax, %r11 ; \
|
|
120
|
+
adoxq %rbx, %r12 ; \
|
|
121
|
+
mulxq 0x18+P1, %rax, %r13 ; \
|
|
122
|
+
adcxq %rax, %r12 ; \
|
|
123
|
+
adoxq %rcx, %r13 ; \
|
|
124
|
+
adcxq %rcx, %r13 ; \
|
|
125
|
+
xorl %ecx, %ecx ; \
|
|
126
|
+
movq 0x10+P2, %rdx ; \
|
|
127
|
+
mulxq P1, %rax, %rbx ; \
|
|
128
|
+
adcxq %rax, %r10 ; \
|
|
129
|
+
adoxq %rbx, %r11 ; \
|
|
130
|
+
mulxq 0x8+P1, %rax, %rbx ; \
|
|
131
|
+
adcxq %rax, %r11 ; \
|
|
132
|
+
adoxq %rbx, %r12 ; \
|
|
133
|
+
mulxq 0x10+P1, %rax, %rbx ; \
|
|
134
|
+
adcxq %rax, %r12 ; \
|
|
135
|
+
adoxq %rbx, %r13 ; \
|
|
136
|
+
mulxq 0x18+P1, %rax, %r14 ; \
|
|
137
|
+
adcxq %rax, %r13 ; \
|
|
138
|
+
adoxq %rcx, %r14 ; \
|
|
139
|
+
adcxq %rcx, %r14 ; \
|
|
140
|
+
xorl %ecx, %ecx ; \
|
|
141
|
+
movq 0x18+P2, %rdx ; \
|
|
142
|
+
mulxq P1, %rax, %rbx ; \
|
|
143
|
+
adcxq %rax, %r11 ; \
|
|
144
|
+
adoxq %rbx, %r12 ; \
|
|
145
|
+
mulxq 0x8+P1, %rax, %rbx ; \
|
|
146
|
+
adcxq %rax, %r12 ; \
|
|
147
|
+
adoxq %rbx, %r13 ; \
|
|
148
|
+
mulxq 0x10+P1, %rax, %rbx ; \
|
|
149
|
+
adcxq %rax, %r13 ; \
|
|
150
|
+
adoxq %rbx, %r14 ; \
|
|
151
|
+
mulxq 0x18+P1, %rax, %r15 ; \
|
|
152
|
+
adcxq %rax, %r14 ; \
|
|
153
|
+
adoxq %rcx, %r15 ; \
|
|
154
|
+
adcxq %rcx, %r15 ; \
|
|
155
|
+
movl $0x26, %edx ; \
|
|
156
|
+
xorl %ecx, %ecx ; \
|
|
157
|
+
mulxq %r12, %rax, %rbx ; \
|
|
158
|
+
adcxq %rax, %r8 ; \
|
|
159
|
+
adoxq %rbx, %r9 ; \
|
|
160
|
+
mulxq %r13, %rax, %rbx ; \
|
|
161
|
+
adcxq %rax, %r9 ; \
|
|
162
|
+
adoxq %rbx, %r10 ; \
|
|
163
|
+
mulxq %r14, %rax, %rbx ; \
|
|
164
|
+
adcxq %rax, %r10 ; \
|
|
165
|
+
adoxq %rbx, %r11 ; \
|
|
166
|
+
mulxq %r15, %rax, %r12 ; \
|
|
167
|
+
adcxq %rax, %r11 ; \
|
|
168
|
+
adoxq %rcx, %r12 ; \
|
|
169
|
+
adcxq %rcx, %r12 ; \
|
|
170
|
+
shldq $0x1, %r11, %r12 ; \
|
|
171
|
+
movl $0x13, %edx ; \
|
|
172
|
+
incq %r12; \
|
|
173
|
+
bts $63, %r11 ; \
|
|
174
|
+
mulxq %r12, %rax, %rbx ; \
|
|
175
|
+
addq %rax, %r8 ; \
|
|
176
|
+
adcq %rbx, %r9 ; \
|
|
177
|
+
adcq %rcx, %r10 ; \
|
|
178
|
+
adcq %rcx, %r11 ; \
|
|
179
|
+
sbbq %rax, %rax ; \
|
|
180
|
+
notq %rax; \
|
|
181
|
+
andq %rdx, %rax ; \
|
|
182
|
+
subq %rax, %r8 ; \
|
|
183
|
+
sbbq %rcx, %r9 ; \
|
|
184
|
+
sbbq %rcx, %r10 ; \
|
|
185
|
+
sbbq %rcx, %r11 ; \
|
|
186
|
+
btr $63, %r11 ; \
|
|
187
|
+
movq %r8, P0 ; \
|
|
188
|
+
movq %r9, 0x8+P0 ; \
|
|
189
|
+
movq %r10, 0x10+P0 ; \
|
|
190
|
+
movq %r11, 0x18+P0
|
|
191
|
+
|
|
192
|
+
// A version of multiplication that only guarantees output < 2 * p_25519.
|
|
193
|
+
// This basically skips the +1 and final correction in quotient estimation.
|
|
194
|
+
|
|
195
|
+
#define mul_4(P0,P1,P2) \
|
|
196
|
+
xorl %ecx, %ecx ; \
|
|
197
|
+
movq P2, %rdx ; \
|
|
198
|
+
mulxq P1, %r8, %r9 ; \
|
|
199
|
+
mulxq 0x8+P1, %rax, %r10 ; \
|
|
200
|
+
addq %rax, %r9 ; \
|
|
201
|
+
mulxq 0x10+P1, %rax, %r11 ; \
|
|
202
|
+
adcq %rax, %r10 ; \
|
|
203
|
+
mulxq 0x18+P1, %rax, %r12 ; \
|
|
204
|
+
adcq %rax, %r11 ; \
|
|
205
|
+
adcq %rcx, %r12 ; \
|
|
206
|
+
xorl %ecx, %ecx ; \
|
|
207
|
+
movq 0x8+P2, %rdx ; \
|
|
208
|
+
mulxq P1, %rax, %rbx ; \
|
|
209
|
+
adcxq %rax, %r9 ; \
|
|
210
|
+
adoxq %rbx, %r10 ; \
|
|
211
|
+
mulxq 0x8+P1, %rax, %rbx ; \
|
|
212
|
+
adcxq %rax, %r10 ; \
|
|
213
|
+
adoxq %rbx, %r11 ; \
|
|
214
|
+
mulxq 0x10+P1, %rax, %rbx ; \
|
|
215
|
+
adcxq %rax, %r11 ; \
|
|
216
|
+
adoxq %rbx, %r12 ; \
|
|
217
|
+
mulxq 0x18+P1, %rax, %r13 ; \
|
|
218
|
+
adcxq %rax, %r12 ; \
|
|
219
|
+
adoxq %rcx, %r13 ; \
|
|
220
|
+
adcxq %rcx, %r13 ; \
|
|
221
|
+
xorl %ecx, %ecx ; \
|
|
222
|
+
movq 0x10+P2, %rdx ; \
|
|
223
|
+
mulxq P1, %rax, %rbx ; \
|
|
224
|
+
adcxq %rax, %r10 ; \
|
|
225
|
+
adoxq %rbx, %r11 ; \
|
|
226
|
+
mulxq 0x8+P1, %rax, %rbx ; \
|
|
227
|
+
adcxq %rax, %r11 ; \
|
|
228
|
+
adoxq %rbx, %r12 ; \
|
|
229
|
+
mulxq 0x10+P1, %rax, %rbx ; \
|
|
230
|
+
adcxq %rax, %r12 ; \
|
|
231
|
+
adoxq %rbx, %r13 ; \
|
|
232
|
+
mulxq 0x18+P1, %rax, %r14 ; \
|
|
233
|
+
adcxq %rax, %r13 ; \
|
|
234
|
+
adoxq %rcx, %r14 ; \
|
|
235
|
+
adcxq %rcx, %r14 ; \
|
|
236
|
+
xorl %ecx, %ecx ; \
|
|
237
|
+
movq 0x18+P2, %rdx ; \
|
|
238
|
+
mulxq P1, %rax, %rbx ; \
|
|
239
|
+
adcxq %rax, %r11 ; \
|
|
240
|
+
adoxq %rbx, %r12 ; \
|
|
241
|
+
mulxq 0x8+P1, %rax, %rbx ; \
|
|
242
|
+
adcxq %rax, %r12 ; \
|
|
243
|
+
adoxq %rbx, %r13 ; \
|
|
244
|
+
mulxq 0x10+P1, %rax, %rbx ; \
|
|
245
|
+
adcxq %rax, %r13 ; \
|
|
246
|
+
adoxq %rbx, %r14 ; \
|
|
247
|
+
mulxq 0x18+P1, %rax, %r15 ; \
|
|
248
|
+
adcxq %rax, %r14 ; \
|
|
249
|
+
adoxq %rcx, %r15 ; \
|
|
250
|
+
adcxq %rcx, %r15 ; \
|
|
251
|
+
movl $0x26, %edx ; \
|
|
252
|
+
xorl %ecx, %ecx ; \
|
|
253
|
+
mulxq %r12, %rax, %rbx ; \
|
|
254
|
+
adcxq %rax, %r8 ; \
|
|
255
|
+
adoxq %rbx, %r9 ; \
|
|
256
|
+
mulxq %r13, %rax, %rbx ; \
|
|
257
|
+
adcxq %rax, %r9 ; \
|
|
258
|
+
adoxq %rbx, %r10 ; \
|
|
259
|
+
mulxq %r14, %rax, %rbx ; \
|
|
260
|
+
adcxq %rax, %r10 ; \
|
|
261
|
+
adoxq %rbx, %r11 ; \
|
|
262
|
+
mulxq %r15, %rax, %r12 ; \
|
|
263
|
+
adcxq %rax, %r11 ; \
|
|
264
|
+
adoxq %rcx, %r12 ; \
|
|
265
|
+
adcxq %rcx, %r12 ; \
|
|
266
|
+
shldq $0x1, %r11, %r12 ; \
|
|
267
|
+
btr $0x3f, %r11 ; \
|
|
268
|
+
movl $0x13, %edx ; \
|
|
269
|
+
imulq %r12, %rdx ; \
|
|
270
|
+
addq %rdx, %r8 ; \
|
|
271
|
+
adcq %rcx, %r9 ; \
|
|
272
|
+
adcq %rcx, %r10 ; \
|
|
273
|
+
adcq %rcx, %r11 ; \
|
|
274
|
+
movq %r8, P0 ; \
|
|
275
|
+
movq %r9, 0x8+P0 ; \
|
|
276
|
+
movq %r10, 0x10+P0 ; \
|
|
277
|
+
movq %r11, 0x18+P0
|
|
278
|
+
|
|
279
|
+
// Squaring just giving a result < 2 * p_25519, which is done by
|
|
280
|
+
// basically skipping the +1 in the quotient estimate and the final
|
|
281
|
+
// optional correction.
|
|
282
|
+
|
|
283
|
+
#define sqr_4(P0,P1) \
|
|
284
|
+
movq P1, %rdx ; \
|
|
285
|
+
mulxq %rdx, %r8, %r15 ; \
|
|
286
|
+
mulxq 0x8+P1, %r9, %r10 ; \
|
|
287
|
+
mulxq 0x18+P1, %r11, %r12 ; \
|
|
288
|
+
movq 0x10+P1, %rdx ; \
|
|
289
|
+
mulxq 0x18+P1, %r13, %r14 ; \
|
|
290
|
+
xorl %ebx, %ebx ; \
|
|
291
|
+
mulxq P1, %rax, %rcx ; \
|
|
292
|
+
adcxq %rax, %r10 ; \
|
|
293
|
+
adoxq %rcx, %r11 ; \
|
|
294
|
+
mulxq 0x8+P1, %rax, %rcx ; \
|
|
295
|
+
adcxq %rax, %r11 ; \
|
|
296
|
+
adoxq %rcx, %r12 ; \
|
|
297
|
+
movq 0x18+P1, %rdx ; \
|
|
298
|
+
mulxq 0x8+P1, %rax, %rcx ; \
|
|
299
|
+
adcxq %rax, %r12 ; \
|
|
300
|
+
adoxq %rcx, %r13 ; \
|
|
301
|
+
adcxq %rbx, %r13 ; \
|
|
302
|
+
adoxq %rbx, %r14 ; \
|
|
303
|
+
adcq %rbx, %r14 ; \
|
|
304
|
+
xorl %ebx, %ebx ; \
|
|
305
|
+
adcxq %r9, %r9 ; \
|
|
306
|
+
adoxq %r15, %r9 ; \
|
|
307
|
+
movq 0x8+P1, %rdx ; \
|
|
308
|
+
mulxq %rdx, %rax, %rdx ; \
|
|
309
|
+
adcxq %r10, %r10 ; \
|
|
310
|
+
adoxq %rax, %r10 ; \
|
|
311
|
+
adcxq %r11, %r11 ; \
|
|
312
|
+
adoxq %rdx, %r11 ; \
|
|
313
|
+
movq 0x10+P1, %rdx ; \
|
|
314
|
+
mulxq %rdx, %rax, %rdx ; \
|
|
315
|
+
adcxq %r12, %r12 ; \
|
|
316
|
+
adoxq %rax, %r12 ; \
|
|
317
|
+
adcxq %r13, %r13 ; \
|
|
318
|
+
adoxq %rdx, %r13 ; \
|
|
319
|
+
movq 0x18+P1, %rdx ; \
|
|
320
|
+
mulxq %rdx, %rax, %r15 ; \
|
|
321
|
+
adcxq %r14, %r14 ; \
|
|
322
|
+
adoxq %rax, %r14 ; \
|
|
323
|
+
adcxq %rbx, %r15 ; \
|
|
324
|
+
adoxq %rbx, %r15 ; \
|
|
325
|
+
movl $0x26, %edx ; \
|
|
326
|
+
xorl %ebx, %ebx ; \
|
|
327
|
+
mulxq %r12, %rax, %rcx ; \
|
|
328
|
+
adcxq %rax, %r8 ; \
|
|
329
|
+
adoxq %rcx, %r9 ; \
|
|
330
|
+
mulxq %r13, %rax, %rcx ; \
|
|
331
|
+
adcxq %rax, %r9 ; \
|
|
332
|
+
adoxq %rcx, %r10 ; \
|
|
333
|
+
mulxq %r14, %rax, %rcx ; \
|
|
334
|
+
adcxq %rax, %r10 ; \
|
|
335
|
+
adoxq %rcx, %r11 ; \
|
|
336
|
+
mulxq %r15, %rax, %r12 ; \
|
|
337
|
+
adcxq %rax, %r11 ; \
|
|
338
|
+
adoxq %rbx, %r12 ; \
|
|
339
|
+
adcxq %rbx, %r12 ; \
|
|
340
|
+
shldq $0x1, %r11, %r12 ; \
|
|
341
|
+
btr $0x3f, %r11 ; \
|
|
342
|
+
movl $0x13, %edx ; \
|
|
343
|
+
imulq %r12, %rdx ; \
|
|
344
|
+
addq %rdx, %r8 ; \
|
|
345
|
+
adcq %rbx, %r9 ; \
|
|
346
|
+
adcq %rbx, %r10 ; \
|
|
347
|
+
adcq %rbx, %r11 ; \
|
|
348
|
+
movq %r8, P0 ; \
|
|
349
|
+
movq %r9, 0x8+P0 ; \
|
|
350
|
+
movq %r10, 0x10+P0 ; \
|
|
351
|
+
movq %r11, 0x18+P0
|
|
352
|
+
|
|
353
|
+
// Modular subtraction with double modulus 2 * p_25519 = 2^256 - 38
|
|
354
|
+
|
|
355
|
+
#define sub_twice4(P0,P1,P2) \
|
|
356
|
+
movq P1, %r8 ; \
|
|
357
|
+
xorl %ebx, %ebx ; \
|
|
358
|
+
subq P2, %r8 ; \
|
|
359
|
+
movq 8+P1, %r9 ; \
|
|
360
|
+
sbbq 8+P2, %r9 ; \
|
|
361
|
+
movl $38, %ecx ; \
|
|
362
|
+
movq 16+P1, %r10 ; \
|
|
363
|
+
sbbq 16+P2, %r10 ; \
|
|
364
|
+
movq 24+P1, %rax ; \
|
|
365
|
+
sbbq 24+P2, %rax ; \
|
|
366
|
+
cmovncq %rbx, %rcx ; \
|
|
367
|
+
subq %rcx, %r8 ; \
|
|
368
|
+
sbbq %rbx, %r9 ; \
|
|
369
|
+
sbbq %rbx, %r10 ; \
|
|
370
|
+
sbbq %rbx, %rax ; \
|
|
371
|
+
movq %r8, P0 ; \
|
|
372
|
+
movq %r9, 8+P0 ; \
|
|
373
|
+
movq %r10, 16+P0 ; \
|
|
374
|
+
movq %rax, 24+P0
|
|
375
|
+
|
|
376
|
+
// Modular addition and doubling with double modulus 2 * p_25519 = 2^256 - 38.
|
|
377
|
+
// This only ensures that the result fits in 4 digits, not that it is reduced
|
|
378
|
+
// even w.r.t. double modulus. The result is always correct modulo provided
|
|
379
|
+
// the sum of the inputs is < 2^256 + 2^256 - 38, so in particular provided
|
|
380
|
+
// at least one of them is reduced double modulo.
|
|
381
|
+
|
|
382
|
+
#define add_twice4(P0,P1,P2) \
|
|
383
|
+
movq P1, %r8 ; \
|
|
384
|
+
xorl %ecx, %ecx ; \
|
|
385
|
+
addq P2, %r8 ; \
|
|
386
|
+
movq 0x8+P1, %r9 ; \
|
|
387
|
+
adcq 0x8+P2, %r9 ; \
|
|
388
|
+
movq 0x10+P1, %r10 ; \
|
|
389
|
+
adcq 0x10+P2, %r10 ; \
|
|
390
|
+
movq 0x18+P1, %r11 ; \
|
|
391
|
+
adcq 0x18+P2, %r11 ; \
|
|
392
|
+
movl $38, %eax ; \
|
|
393
|
+
cmovncq %rcx, %rax ; \
|
|
394
|
+
addq %rax, %r8 ; \
|
|
395
|
+
adcq %rcx, %r9 ; \
|
|
396
|
+
adcq %rcx, %r10 ; \
|
|
397
|
+
adcq %rcx, %r11 ; \
|
|
398
|
+
movq %r8, P0 ; \
|
|
399
|
+
movq %r9, 0x8+P0 ; \
|
|
400
|
+
movq %r10, 0x10+P0 ; \
|
|
401
|
+
movq %r11, 0x18+P0
|
|
402
|
+
|
|
403
|
+
#define double_twice4(P0,P1) \
|
|
404
|
+
movq P1, %r8 ; \
|
|
405
|
+
xorl %ecx, %ecx ; \
|
|
406
|
+
addq %r8, %r8 ; \
|
|
407
|
+
movq 0x8+P1, %r9 ; \
|
|
408
|
+
adcq %r9, %r9 ; \
|
|
409
|
+
movq 0x10+P1, %r10 ; \
|
|
410
|
+
adcq %r10, %r10 ; \
|
|
411
|
+
movq 0x18+P1, %r11 ; \
|
|
412
|
+
adcq %r11, %r11 ; \
|
|
413
|
+
movl $38, %eax ; \
|
|
414
|
+
cmovncq %rcx, %rax ; \
|
|
415
|
+
addq %rax, %r8 ; \
|
|
416
|
+
adcq %rcx, %r9 ; \
|
|
417
|
+
adcq %rcx, %r10 ; \
|
|
418
|
+
adcq %rcx, %r11 ; \
|
|
419
|
+
movq %r8, P0 ; \
|
|
420
|
+
movq %r9, 0x8+P0 ; \
|
|
421
|
+
movq %r10, 0x10+P0 ; \
|
|
422
|
+
movq %r11, 0x18+P0
|
|
423
|
+
|
|
424
|
+
// Load the constant k_25519 = 2 * d_25519 using immediate operations
|
|
425
|
+
|
|
426
|
+
#define load_k25519(P0) \
|
|
427
|
+
movq $0xebd69b9426b2f159, %rax ; \
|
|
428
|
+
movq %rax, P0 ; \
|
|
429
|
+
movq $0x00e0149a8283b156, %rax ; \
|
|
430
|
+
movq %rax, 8+P0 ; \
|
|
431
|
+
movq $0x198e80f2eef3d130, %rax ; \
|
|
432
|
+
movq %rax, 16+P0 ; \
|
|
433
|
+
movq $0x2406d9dc56dffce7, %rax ; \
|
|
434
|
+
movq %rax, 24+P0
|
|
435
|
+
|
|
436
|
+
S2N_BN_SYMBOL(edwards25519_scalarmuldouble):
|
|
437
|
+
|
|
438
|
+
// In this case the Windows form literally makes a subroutine call.
|
|
439
|
+
// This avoids hassle arising from keeping code and data together.
|
|
440
|
+
|
|
441
|
+
#if WINDOWS_ABI
|
|
442
|
+
pushq %rdi
|
|
443
|
+
pushq %rsi
|
|
444
|
+
movq %rcx, %rdi
|
|
445
|
+
movq %rdx, %rsi
|
|
446
|
+
movq %r8, %rdx
|
|
447
|
+
movq %r9, %rcx
|
|
448
|
+
callq edwards25519_scalarmuldouble_standard
|
|
449
|
+
popq %rsi
|
|
450
|
+
popq %rdi
|
|
451
|
+
ret
|
|
452
|
+
|
|
453
|
+
edwards25519_scalarmuldouble_standard:
|
|
454
|
+
#endif
|
|
455
|
+
|
|
456
|
+
// Save registers, make room for temps, preserve input arguments.
|
|
457
|
+
|
|
458
|
+
pushq %rbx
|
|
459
|
+
pushq %rbp
|
|
460
|
+
pushq %r12
|
|
461
|
+
pushq %r13
|
|
462
|
+
pushq %r14
|
|
463
|
+
pushq %r15
|
|
464
|
+
subq $NSPACE, %rsp
|
|
465
|
+
|
|
466
|
+
// Move the output pointer to a stable place
|
|
467
|
+
|
|
468
|
+
movq %rdi, res
|
|
469
|
+
|
|
470
|
+
// Copy scalars while recoding all 4-bit nybbles except the top
|
|
471
|
+
// one (bits 252..255) into signed 4-bit digits. This is essentially
|
|
472
|
+
// done just by adding the recoding constant 0x0888..888, after
|
|
473
|
+
// which all digits except the first have an implicit bias of -8,
|
|
474
|
+
// so 0 -> -8, 1 -> -7, ... 7 -> -1, 8 -> 0, 9 -> 1, ... 15 -> 7.
|
|
475
|
+
// (We could literally create 2s complement signed nybbles by
|
|
476
|
+
// XORing with the same constant 0x0888..888 afterwards, but it
|
|
477
|
+
// doesn't seem to make the end usage any simpler.)
|
|
478
|
+
//
|
|
479
|
+
// In order to ensure that the unrecoded top nybble (bits 252..255)
|
|
480
|
+
// does not become > 8 as a result of carries lower down from the
|
|
481
|
+
// recoding, we first (conceptually) subtract the group order iff
|
|
482
|
+
// the top digit of the scalar is > 2^63. In the implementation the
|
|
483
|
+
// reduction and recoding are combined by optionally using the
|
|
484
|
+
// modified recoding constant 0x0888...888 + (2^256 - group_order).
|
|
485
|
+
|
|
486
|
+
movq (%rcx), %r8
|
|
487
|
+
movq 8(%rcx), %r9
|
|
488
|
+
movq 16(%rcx), %r10
|
|
489
|
+
movq 24(%rcx), %r11
|
|
490
|
+
movq $0xc7f56fb5a0d9e920, %r12
|
|
491
|
+
movq $0xe190b99370cba1d5, %r13
|
|
492
|
+
movq $0x8888888888888887, %r14
|
|
493
|
+
movq $0x8888888888888888, %r15
|
|
494
|
+
movq $0x8000000000000000, %rax
|
|
495
|
+
movq $0x0888888888888888, %rbx
|
|
496
|
+
cmpq %r11, %rax
|
|
497
|
+
cmovncq %r15, %r12
|
|
498
|
+
cmovncq %r15, %r13
|
|
499
|
+
cmovncq %r15, %r14
|
|
500
|
+
cmovncq %rbx, %r15
|
|
501
|
+
addq %r12, %r8
|
|
502
|
+
adcq %r13, %r9
|
|
503
|
+
adcq %r14, %r10
|
|
504
|
+
adcq %r15, %r11
|
|
505
|
+
movq %r8, BSCALAR(%rsp)
|
|
506
|
+
movq %r9, BSCALAR+8(%rsp)
|
|
507
|
+
movq %r10, BSCALAR+16(%rsp)
|
|
508
|
+
movq %r11, BSCALAR+24(%rsp)
|
|
509
|
+
|
|
510
|
+
movq (%rsi), %r8
|
|
511
|
+
movq 8(%rsi), %r9
|
|
512
|
+
movq 16(%rsi), %r10
|
|
513
|
+
movq 24(%rsi), %r11
|
|
514
|
+
movq $0xc7f56fb5a0d9e920, %r12
|
|
515
|
+
movq $0xe190b99370cba1d5, %r13
|
|
516
|
+
movq $0x8888888888888887, %r14
|
|
517
|
+
movq $0x8888888888888888, %r15
|
|
518
|
+
movq $0x8000000000000000, %rax
|
|
519
|
+
movq $0x0888888888888888, %rbx
|
|
520
|
+
cmpq %r11, %rax
|
|
521
|
+
cmovncq %r15, %r12
|
|
522
|
+
cmovncq %r15, %r13
|
|
523
|
+
cmovncq %r15, %r14
|
|
524
|
+
cmovncq %rbx, %r15
|
|
525
|
+
addq %r12, %r8
|
|
526
|
+
adcq %r13, %r9
|
|
527
|
+
adcq %r14, %r10
|
|
528
|
+
adcq %r15, %r11
|
|
529
|
+
movq %r8, SCALAR(%rsp)
|
|
530
|
+
movq %r9, SCALAR+8(%rsp)
|
|
531
|
+
movq %r10, SCALAR+16(%rsp)
|
|
532
|
+
movq %r11, SCALAR+24(%rsp)
|
|
533
|
+
|
|
534
|
+
// Create table of multiples 1..8 of the general input point at "tab".
|
|
535
|
+
// Reduce the input coordinates x and y modulo 2^256 - 38 first, for the
|
|
536
|
+
// sake of definiteness; this is the reduction that will be maintained.
|
|
537
|
+
// We could slightly optimize the additions because we know the input
|
|
538
|
+
// point is affine (so Z = 1), but it doesn't seem worth the complication.
|
|
539
|
+
|
|
540
|
+
movl $38, %eax
|
|
541
|
+
movq (%rdx), %r8
|
|
542
|
+
xorl %ebx, %ebx
|
|
543
|
+
movq 8(%rdx), %r9
|
|
544
|
+
xorl %ecx, %ecx
|
|
545
|
+
movq 16(%rdx), %r10
|
|
546
|
+
xorl %esi, %esi
|
|
547
|
+
movq 24(%rdx), %r11
|
|
548
|
+
addq %r8, %rax
|
|
549
|
+
adcq %r9, %rbx
|
|
550
|
+
adcq %r10, %rcx
|
|
551
|
+
adcq %r11, %rsi
|
|
552
|
+
cmovncq %r8, %rax
|
|
553
|
+
movq %rax, TAB(%rsp)
|
|
554
|
+
cmovncq %r9, %rbx
|
|
555
|
+
movq %rbx, TAB+8(%rsp)
|
|
556
|
+
cmovncq %r10, %rcx
|
|
557
|
+
movq %rcx, TAB+16(%rsp)
|
|
558
|
+
cmovncq %r11, %rsi
|
|
559
|
+
movq %rsi, TAB+24(%rsp)
|
|
560
|
+
|
|
561
|
+
movl $38, %eax
|
|
562
|
+
movq 32(%rdx), %r8
|
|
563
|
+
xorl %ebx, %ebx
|
|
564
|
+
movq 40(%rdx), %r9
|
|
565
|
+
xorl %ecx, %ecx
|
|
566
|
+
movq 48(%rdx), %r10
|
|
567
|
+
xorl %esi, %esi
|
|
568
|
+
movq 56(%rdx), %r11
|
|
569
|
+
addq %r8, %rax
|
|
570
|
+
adcq %r9, %rbx
|
|
571
|
+
adcq %r10, %rcx
|
|
572
|
+
adcq %r11, %rsi
|
|
573
|
+
cmovncq %r8, %rax
|
|
574
|
+
movq %rax, TAB+32(%rsp)
|
|
575
|
+
cmovncq %r9, %rbx
|
|
576
|
+
movq %rbx, TAB+40(%rsp)
|
|
577
|
+
cmovncq %r10, %rcx
|
|
578
|
+
movq %rcx, TAB+48(%rsp)
|
|
579
|
+
cmovncq %r11, %rsi
|
|
580
|
+
movq %rsi, TAB+56(%rsp)
|
|
581
|
+
|
|
582
|
+
movl $1, %eax
|
|
583
|
+
movq %rax, TAB+64(%rsp)
|
|
584
|
+
xorl %eax, %eax
|
|
585
|
+
movq %rax, TAB+72(%rsp)
|
|
586
|
+
movq %rax, TAB+80(%rsp)
|
|
587
|
+
movq %rax, TAB+88(%rsp)
|
|
588
|
+
|
|
589
|
+
leaq TAB+96(%rsp), %rdi
|
|
590
|
+
leaq TAB(%rsp), %rsi
|
|
591
|
+
leaq TAB+32(%rsp), %rbp
|
|
592
|
+
mul_4(x_0,x_1,x_2)
|
|
593
|
+
|
|
594
|
+
// Multiple 2
|
|
595
|
+
|
|
596
|
+
leaq TAB+1*128(%rsp), %rdi
|
|
597
|
+
leaq TAB(%rsp), %rsi
|
|
598
|
+
callq edwards25519_scalarmuldouble_epdouble
|
|
599
|
+
|
|
600
|
+
// Multiple 3
|
|
601
|
+
|
|
602
|
+
leaq TAB+2*128(%rsp), %rdi
|
|
603
|
+
leaq TAB(%rsp), %rsi
|
|
604
|
+
leaq TAB+1*128(%rsp), %rbp
|
|
605
|
+
callq edwards25519_scalarmuldouble_epadd
|
|
606
|
+
|
|
607
|
+
// Multiple 4
|
|
608
|
+
|
|
609
|
+
leaq TAB+3*128(%rsp), %rdi
|
|
610
|
+
leaq TAB+1*128(%rsp), %rsi
|
|
611
|
+
callq edwards25519_scalarmuldouble_epdouble
|
|
612
|
+
|
|
613
|
+
// Multiple 5
|
|
614
|
+
|
|
615
|
+
leaq TAB+4*128(%rsp), %rdi
|
|
616
|
+
leaq TAB(%rsp), %rsi
|
|
617
|
+
leaq TAB+3*128(%rsp), %rbp
|
|
618
|
+
callq edwards25519_scalarmuldouble_epadd
|
|
619
|
+
|
|
620
|
+
// Multiple 6
|
|
621
|
+
|
|
622
|
+
leaq TAB+5*128(%rsp), %rdi
|
|
623
|
+
leaq TAB+2*128(%rsp), %rsi
|
|
624
|
+
callq edwards25519_scalarmuldouble_epdouble
|
|
625
|
+
|
|
626
|
+
// Multiple 7
|
|
627
|
+
|
|
628
|
+
leaq TAB+6*128(%rsp), %rdi
|
|
629
|
+
leaq TAB(%rsp), %rsi
|
|
630
|
+
leaq TAB+5*128(%rsp), %rbp
|
|
631
|
+
callq edwards25519_scalarmuldouble_epadd
|
|
632
|
+
|
|
633
|
+
// Multiple 8
|
|
634
|
+
|
|
635
|
+
leaq TAB+7*128(%rsp), %rdi
|
|
636
|
+
leaq TAB+3*128(%rsp), %rsi
|
|
637
|
+
callq edwards25519_scalarmuldouble_epdouble
|
|
638
|
+
|
|
639
|
+
// Handle the initialization, starting the loop counter at i = 252
|
|
640
|
+
// and initializing acc to the sum of the table entries for the
|
|
641
|
+
// top nybbles of the scalars (the ones with no implicit -8 bias).
|
|
642
|
+
|
|
643
|
+
movq $252, %rax
|
|
644
|
+
movq %rax, i
|
|
645
|
+
|
|
646
|
+
// Index for btable entry...
|
|
647
|
+
|
|
648
|
+
movq BSCALAR+24(%rsp), %rax
|
|
649
|
+
shrq $60, %rax
|
|
650
|
+
movq %rax, bf
|
|
651
|
+
|
|
652
|
+
// ...and constant-time indexing based on that index
|
|
653
|
+
|
|
654
|
+
movl $1, %eax
|
|
655
|
+
xorl %ebx, %ebx
|
|
656
|
+
xorl %ecx, %ecx
|
|
657
|
+
xorl %edx, %edx
|
|
658
|
+
movl $1, %r8d
|
|
659
|
+
xorl %r9d, %r9d
|
|
660
|
+
xorl %r10d, %r10d
|
|
661
|
+
xorl %r11d, %r11d
|
|
662
|
+
xorl %r12d, %r12d
|
|
663
|
+
xorl %r13d, %r13d
|
|
664
|
+
xorl %r14d, %r14d
|
|
665
|
+
xorl %r15d, %r15d
|
|
666
|
+
|
|
667
|
+
leaq edwards25519_scalarmuldouble_table(%rip), %rbp
|
|
668
|
+
|
|
669
|
+
cmpq $1, bf
|
|
670
|
+
movq (%rbp), %rsi
|
|
671
|
+
cmovzq %rsi, %rax
|
|
672
|
+
movq 8(%rbp), %rsi
|
|
673
|
+
cmovzq %rsi, %rbx
|
|
674
|
+
movq 16(%rbp), %rsi
|
|
675
|
+
cmovzq %rsi, %rcx
|
|
676
|
+
movq 24(%rbp), %rsi
|
|
677
|
+
cmovzq %rsi, %rdx
|
|
678
|
+
movq 32(%rbp), %rsi
|
|
679
|
+
cmovzq %rsi, %r8
|
|
680
|
+
movq 40(%rbp), %rsi
|
|
681
|
+
cmovzq %rsi, %r9
|
|
682
|
+
movq 48(%rbp), %rsi
|
|
683
|
+
cmovzq %rsi, %r10
|
|
684
|
+
movq 56(%rbp), %rsi
|
|
685
|
+
cmovzq %rsi, %r11
|
|
686
|
+
movq 64(%rbp), %rsi
|
|
687
|
+
cmovzq %rsi, %r12
|
|
688
|
+
movq 72(%rbp), %rsi
|
|
689
|
+
cmovzq %rsi, %r13
|
|
690
|
+
movq 80(%rbp), %rsi
|
|
691
|
+
cmovzq %rsi, %r14
|
|
692
|
+
movq 88(%rbp), %rsi
|
|
693
|
+
cmovzq %rsi, %r15
|
|
694
|
+
addq $96, %rbp
|
|
695
|
+
|
|
696
|
+
cmpq $2, bf
|
|
697
|
+
movq (%rbp), %rsi
|
|
698
|
+
cmovzq %rsi, %rax
|
|
699
|
+
movq 8(%rbp), %rsi
|
|
700
|
+
cmovzq %rsi, %rbx
|
|
701
|
+
movq 16(%rbp), %rsi
|
|
702
|
+
cmovzq %rsi, %rcx
|
|
703
|
+
movq 24(%rbp), %rsi
|
|
704
|
+
cmovzq %rsi, %rdx
|
|
705
|
+
movq 32(%rbp), %rsi
|
|
706
|
+
cmovzq %rsi, %r8
|
|
707
|
+
movq 40(%rbp), %rsi
|
|
708
|
+
cmovzq %rsi, %r9
|
|
709
|
+
movq 48(%rbp), %rsi
|
|
710
|
+
cmovzq %rsi, %r10
|
|
711
|
+
movq 56(%rbp), %rsi
|
|
712
|
+
cmovzq %rsi, %r11
|
|
713
|
+
movq 64(%rbp), %rsi
|
|
714
|
+
cmovzq %rsi, %r12
|
|
715
|
+
movq 72(%rbp), %rsi
|
|
716
|
+
cmovzq %rsi, %r13
|
|
717
|
+
movq 80(%rbp), %rsi
|
|
718
|
+
cmovzq %rsi, %r14
|
|
719
|
+
movq 88(%rbp), %rsi
|
|
720
|
+
cmovzq %rsi, %r15
|
|
721
|
+
addq $96, %rbp
|
|
722
|
+
|
|
723
|
+
cmpq $3, bf
|
|
724
|
+
movq (%rbp), %rsi
|
|
725
|
+
cmovzq %rsi, %rax
|
|
726
|
+
movq 8(%rbp), %rsi
|
|
727
|
+
cmovzq %rsi, %rbx
|
|
728
|
+
movq 16(%rbp), %rsi
|
|
729
|
+
cmovzq %rsi, %rcx
|
|
730
|
+
movq 24(%rbp), %rsi
|
|
731
|
+
cmovzq %rsi, %rdx
|
|
732
|
+
movq 32(%rbp), %rsi
|
|
733
|
+
cmovzq %rsi, %r8
|
|
734
|
+
movq 40(%rbp), %rsi
|
|
735
|
+
cmovzq %rsi, %r9
|
|
736
|
+
movq 48(%rbp), %rsi
|
|
737
|
+
cmovzq %rsi, %r10
|
|
738
|
+
movq 56(%rbp), %rsi
|
|
739
|
+
cmovzq %rsi, %r11
|
|
740
|
+
movq 64(%rbp), %rsi
|
|
741
|
+
cmovzq %rsi, %r12
|
|
742
|
+
movq 72(%rbp), %rsi
|
|
743
|
+
cmovzq %rsi, %r13
|
|
744
|
+
movq 80(%rbp), %rsi
|
|
745
|
+
cmovzq %rsi, %r14
|
|
746
|
+
movq 88(%rbp), %rsi
|
|
747
|
+
cmovzq %rsi, %r15
|
|
748
|
+
addq $96, %rbp
|
|
749
|
+
|
|
750
|
+
cmpq $4, bf
|
|
751
|
+
movq (%rbp), %rsi
|
|
752
|
+
cmovzq %rsi, %rax
|
|
753
|
+
movq 8(%rbp), %rsi
|
|
754
|
+
cmovzq %rsi, %rbx
|
|
755
|
+
movq 16(%rbp), %rsi
|
|
756
|
+
cmovzq %rsi, %rcx
|
|
757
|
+
movq 24(%rbp), %rsi
|
|
758
|
+
cmovzq %rsi, %rdx
|
|
759
|
+
movq 32(%rbp), %rsi
|
|
760
|
+
cmovzq %rsi, %r8
|
|
761
|
+
movq 40(%rbp), %rsi
|
|
762
|
+
cmovzq %rsi, %r9
|
|
763
|
+
movq 48(%rbp), %rsi
|
|
764
|
+
cmovzq %rsi, %r10
|
|
765
|
+
movq 56(%rbp), %rsi
|
|
766
|
+
cmovzq %rsi, %r11
|
|
767
|
+
movq 64(%rbp), %rsi
|
|
768
|
+
cmovzq %rsi, %r12
|
|
769
|
+
movq 72(%rbp), %rsi
|
|
770
|
+
cmovzq %rsi, %r13
|
|
771
|
+
movq 80(%rbp), %rsi
|
|
772
|
+
cmovzq %rsi, %r14
|
|
773
|
+
movq 88(%rbp), %rsi
|
|
774
|
+
cmovzq %rsi, %r15
|
|
775
|
+
addq $96, %rbp
|
|
776
|
+
|
|
777
|
+
cmpq $5, bf
|
|
778
|
+
movq (%rbp), %rsi
|
|
779
|
+
cmovzq %rsi, %rax
|
|
780
|
+
movq 8(%rbp), %rsi
|
|
781
|
+
cmovzq %rsi, %rbx
|
|
782
|
+
movq 16(%rbp), %rsi
|
|
783
|
+
cmovzq %rsi, %rcx
|
|
784
|
+
movq 24(%rbp), %rsi
|
|
785
|
+
cmovzq %rsi, %rdx
|
|
786
|
+
movq 32(%rbp), %rsi
|
|
787
|
+
cmovzq %rsi, %r8
|
|
788
|
+
movq 40(%rbp), %rsi
|
|
789
|
+
cmovzq %rsi, %r9
|
|
790
|
+
movq 48(%rbp), %rsi
|
|
791
|
+
cmovzq %rsi, %r10
|
|
792
|
+
movq 56(%rbp), %rsi
|
|
793
|
+
cmovzq %rsi, %r11
|
|
794
|
+
movq 64(%rbp), %rsi
|
|
795
|
+
cmovzq %rsi, %r12
|
|
796
|
+
movq 72(%rbp), %rsi
|
|
797
|
+
cmovzq %rsi, %r13
|
|
798
|
+
movq 80(%rbp), %rsi
|
|
799
|
+
cmovzq %rsi, %r14
|
|
800
|
+
movq 88(%rbp), %rsi
|
|
801
|
+
cmovzq %rsi, %r15
|
|
802
|
+
addq $96, %rbp
|
|
803
|
+
|
|
804
|
+
cmpq $6, bf
|
|
805
|
+
movq (%rbp), %rsi
|
|
806
|
+
cmovzq %rsi, %rax
|
|
807
|
+
movq 8(%rbp), %rsi
|
|
808
|
+
cmovzq %rsi, %rbx
|
|
809
|
+
movq 16(%rbp), %rsi
|
|
810
|
+
cmovzq %rsi, %rcx
|
|
811
|
+
movq 24(%rbp), %rsi
|
|
812
|
+
cmovzq %rsi, %rdx
|
|
813
|
+
movq 32(%rbp), %rsi
|
|
814
|
+
cmovzq %rsi, %r8
|
|
815
|
+
movq 40(%rbp), %rsi
|
|
816
|
+
cmovzq %rsi, %r9
|
|
817
|
+
movq 48(%rbp), %rsi
|
|
818
|
+
cmovzq %rsi, %r10
|
|
819
|
+
movq 56(%rbp), %rsi
|
|
820
|
+
cmovzq %rsi, %r11
|
|
821
|
+
movq 64(%rbp), %rsi
|
|
822
|
+
cmovzq %rsi, %r12
|
|
823
|
+
movq 72(%rbp), %rsi
|
|
824
|
+
cmovzq %rsi, %r13
|
|
825
|
+
movq 80(%rbp), %rsi
|
|
826
|
+
cmovzq %rsi, %r14
|
|
827
|
+
movq 88(%rbp), %rsi
|
|
828
|
+
cmovzq %rsi, %r15
|
|
829
|
+
addq $96, %rbp
|
|
830
|
+
|
|
831
|
+
cmpq $7, bf
|
|
832
|
+
movq (%rbp), %rsi
|
|
833
|
+
cmovzq %rsi, %rax
|
|
834
|
+
movq 8(%rbp), %rsi
|
|
835
|
+
cmovzq %rsi, %rbx
|
|
836
|
+
movq 16(%rbp), %rsi
|
|
837
|
+
cmovzq %rsi, %rcx
|
|
838
|
+
movq 24(%rbp), %rsi
|
|
839
|
+
cmovzq %rsi, %rdx
|
|
840
|
+
movq 32(%rbp), %rsi
|
|
841
|
+
cmovzq %rsi, %r8
|
|
842
|
+
movq 40(%rbp), %rsi
|
|
843
|
+
cmovzq %rsi, %r9
|
|
844
|
+
movq 48(%rbp), %rsi
|
|
845
|
+
cmovzq %rsi, %r10
|
|
846
|
+
movq 56(%rbp), %rsi
|
|
847
|
+
cmovzq %rsi, %r11
|
|
848
|
+
movq 64(%rbp), %rsi
|
|
849
|
+
cmovzq %rsi, %r12
|
|
850
|
+
movq 72(%rbp), %rsi
|
|
851
|
+
cmovzq %rsi, %r13
|
|
852
|
+
movq 80(%rbp), %rsi
|
|
853
|
+
cmovzq %rsi, %r14
|
|
854
|
+
movq 88(%rbp), %rsi
|
|
855
|
+
cmovzq %rsi, %r15
|
|
856
|
+
addq $96, %rbp
|
|
857
|
+
|
|
858
|
+
cmpq $8, bf
|
|
859
|
+
movq (%rbp), %rsi
|
|
860
|
+
cmovzq %rsi, %rax
|
|
861
|
+
movq 8(%rbp), %rsi
|
|
862
|
+
cmovzq %rsi, %rbx
|
|
863
|
+
movq 16(%rbp), %rsi
|
|
864
|
+
cmovzq %rsi, %rcx
|
|
865
|
+
movq 24(%rbp), %rsi
|
|
866
|
+
cmovzq %rsi, %rdx
|
|
867
|
+
movq 32(%rbp), %rsi
|
|
868
|
+
cmovzq %rsi, %r8
|
|
869
|
+
movq 40(%rbp), %rsi
|
|
870
|
+
cmovzq %rsi, %r9
|
|
871
|
+
movq 48(%rbp), %rsi
|
|
872
|
+
cmovzq %rsi, %r10
|
|
873
|
+
movq 56(%rbp), %rsi
|
|
874
|
+
cmovzq %rsi, %r11
|
|
875
|
+
movq 64(%rbp), %rsi
|
|
876
|
+
cmovzq %rsi, %r12
|
|
877
|
+
movq 72(%rbp), %rsi
|
|
878
|
+
cmovzq %rsi, %r13
|
|
879
|
+
movq 80(%rbp), %rsi
|
|
880
|
+
cmovzq %rsi, %r14
|
|
881
|
+
movq 88(%rbp), %rsi
|
|
882
|
+
cmovzq %rsi, %r15
|
|
883
|
+
|
|
884
|
+
movq %rax, BTABENT(%rsp)
|
|
885
|
+
movq %rbx, BTABENT+8(%rsp)
|
|
886
|
+
movq %rcx, BTABENT+16(%rsp)
|
|
887
|
+
movq %rdx, BTABENT+24(%rsp)
|
|
888
|
+
movq %r8, BTABENT+32(%rsp)
|
|
889
|
+
movq %r9, BTABENT+40(%rsp)
|
|
890
|
+
movq %r10, BTABENT+48(%rsp)
|
|
891
|
+
movq %r11, BTABENT+56(%rsp)
|
|
892
|
+
movq %r12, BTABENT+64(%rsp)
|
|
893
|
+
movq %r13, BTABENT+72(%rsp)
|
|
894
|
+
movq %r14, BTABENT+80(%rsp)
|
|
895
|
+
movq %r15, BTABENT+88(%rsp)
|
|
896
|
+
|
|
897
|
+
// Index for table entry...
|
|
898
|
+
|
|
899
|
+
movq SCALAR+24(%rsp), %rax
|
|
900
|
+
shrq $60, %rax
|
|
901
|
+
movq %rax, bf
|
|
902
|
+
|
|
903
|
+
// ...and constant-time indexing based on that index.
|
|
904
|
+
// Do the Y and Z fields first, to save on registers...
|
|
905
|
+
|
|
906
|
+
movl $1, %eax
|
|
907
|
+
xorl %ebx, %ebx
|
|
908
|
+
xorl %ecx, %ecx
|
|
909
|
+
xorl %edx, %edx
|
|
910
|
+
movl $1, %r8d
|
|
911
|
+
xorl %r9d, %r9d
|
|
912
|
+
xorl %r10d, %r10d
|
|
913
|
+
xorl %r11d, %r11d
|
|
914
|
+
|
|
915
|
+
leaq TAB+32(%rsp), %rbp
|
|
916
|
+
|
|
917
|
+
cmpq $1, bf
|
|
918
|
+
movq (%rbp), %rsi
|
|
919
|
+
cmovzq %rsi, %rax
|
|
920
|
+
movq 8(%rbp), %rsi
|
|
921
|
+
cmovzq %rsi, %rbx
|
|
922
|
+
movq 16(%rbp), %rsi
|
|
923
|
+
cmovzq %rsi, %rcx
|
|
924
|
+
movq 24(%rbp), %rsi
|
|
925
|
+
cmovzq %rsi, %rdx
|
|
926
|
+
movq 32(%rbp), %rsi
|
|
927
|
+
cmovzq %rsi, %r8
|
|
928
|
+
movq 40(%rbp), %rsi
|
|
929
|
+
cmovzq %rsi, %r9
|
|
930
|
+
movq 48(%rbp), %rsi
|
|
931
|
+
cmovzq %rsi, %r10
|
|
932
|
+
movq 56(%rbp), %rsi
|
|
933
|
+
cmovzq %rsi, %r11
|
|
934
|
+
addq $128, %rbp
|
|
935
|
+
|
|
936
|
+
cmpq $2, bf
|
|
937
|
+
movq (%rbp), %rsi
|
|
938
|
+
cmovzq %rsi, %rax
|
|
939
|
+
movq 8(%rbp), %rsi
|
|
940
|
+
cmovzq %rsi, %rbx
|
|
941
|
+
movq 16(%rbp), %rsi
|
|
942
|
+
cmovzq %rsi, %rcx
|
|
943
|
+
movq 24(%rbp), %rsi
|
|
944
|
+
cmovzq %rsi, %rdx
|
|
945
|
+
movq 32(%rbp), %rsi
|
|
946
|
+
cmovzq %rsi, %r8
|
|
947
|
+
movq 40(%rbp), %rsi
|
|
948
|
+
cmovzq %rsi, %r9
|
|
949
|
+
movq 48(%rbp), %rsi
|
|
950
|
+
cmovzq %rsi, %r10
|
|
951
|
+
movq 56(%rbp), %rsi
|
|
952
|
+
cmovzq %rsi, %r11
|
|
953
|
+
addq $128, %rbp
|
|
954
|
+
|
|
955
|
+
cmpq $3, bf
|
|
956
|
+
movq (%rbp), %rsi
|
|
957
|
+
cmovzq %rsi, %rax
|
|
958
|
+
movq 8(%rbp), %rsi
|
|
959
|
+
cmovzq %rsi, %rbx
|
|
960
|
+
movq 16(%rbp), %rsi
|
|
961
|
+
cmovzq %rsi, %rcx
|
|
962
|
+
movq 24(%rbp), %rsi
|
|
963
|
+
cmovzq %rsi, %rdx
|
|
964
|
+
movq 32(%rbp), %rsi
|
|
965
|
+
cmovzq %rsi, %r8
|
|
966
|
+
movq 40(%rbp), %rsi
|
|
967
|
+
cmovzq %rsi, %r9
|
|
968
|
+
movq 48(%rbp), %rsi
|
|
969
|
+
cmovzq %rsi, %r10
|
|
970
|
+
movq 56(%rbp), %rsi
|
|
971
|
+
cmovzq %rsi, %r11
|
|
972
|
+
addq $128, %rbp
|
|
973
|
+
|
|
974
|
+
cmpq $4, bf
|
|
975
|
+
movq (%rbp), %rsi
|
|
976
|
+
cmovzq %rsi, %rax
|
|
977
|
+
movq 8(%rbp), %rsi
|
|
978
|
+
cmovzq %rsi, %rbx
|
|
979
|
+
movq 16(%rbp), %rsi
|
|
980
|
+
cmovzq %rsi, %rcx
|
|
981
|
+
movq 24(%rbp), %rsi
|
|
982
|
+
cmovzq %rsi, %rdx
|
|
983
|
+
movq 32(%rbp), %rsi
|
|
984
|
+
cmovzq %rsi, %r8
|
|
985
|
+
movq 40(%rbp), %rsi
|
|
986
|
+
cmovzq %rsi, %r9
|
|
987
|
+
movq 48(%rbp), %rsi
|
|
988
|
+
cmovzq %rsi, %r10
|
|
989
|
+
movq 56(%rbp), %rsi
|
|
990
|
+
cmovzq %rsi, %r11
|
|
991
|
+
addq $128, %rbp
|
|
992
|
+
|
|
993
|
+
cmpq $5, bf
|
|
994
|
+
movq (%rbp), %rsi
|
|
995
|
+
cmovzq %rsi, %rax
|
|
996
|
+
movq 8(%rbp), %rsi
|
|
997
|
+
cmovzq %rsi, %rbx
|
|
998
|
+
movq 16(%rbp), %rsi
|
|
999
|
+
cmovzq %rsi, %rcx
|
|
1000
|
+
movq 24(%rbp), %rsi
|
|
1001
|
+
cmovzq %rsi, %rdx
|
|
1002
|
+
movq 32(%rbp), %rsi
|
|
1003
|
+
cmovzq %rsi, %r8
|
|
1004
|
+
movq 40(%rbp), %rsi
|
|
1005
|
+
cmovzq %rsi, %r9
|
|
1006
|
+
movq 48(%rbp), %rsi
|
|
1007
|
+
cmovzq %rsi, %r10
|
|
1008
|
+
movq 56(%rbp), %rsi
|
|
1009
|
+
cmovzq %rsi, %r11
|
|
1010
|
+
addq $128, %rbp
|
|
1011
|
+
|
|
1012
|
+
cmpq $6, bf
|
|
1013
|
+
movq (%rbp), %rsi
|
|
1014
|
+
cmovzq %rsi, %rax
|
|
1015
|
+
movq 8(%rbp), %rsi
|
|
1016
|
+
cmovzq %rsi, %rbx
|
|
1017
|
+
movq 16(%rbp), %rsi
|
|
1018
|
+
cmovzq %rsi, %rcx
|
|
1019
|
+
movq 24(%rbp), %rsi
|
|
1020
|
+
cmovzq %rsi, %rdx
|
|
1021
|
+
movq 32(%rbp), %rsi
|
|
1022
|
+
cmovzq %rsi, %r8
|
|
1023
|
+
movq 40(%rbp), %rsi
|
|
1024
|
+
cmovzq %rsi, %r9
|
|
1025
|
+
movq 48(%rbp), %rsi
|
|
1026
|
+
cmovzq %rsi, %r10
|
|
1027
|
+
movq 56(%rbp), %rsi
|
|
1028
|
+
cmovzq %rsi, %r11
|
|
1029
|
+
addq $128, %rbp
|
|
1030
|
+
|
|
1031
|
+
cmpq $7, bf
|
|
1032
|
+
movq (%rbp), %rsi
|
|
1033
|
+
cmovzq %rsi, %rax
|
|
1034
|
+
movq 8(%rbp), %rsi
|
|
1035
|
+
cmovzq %rsi, %rbx
|
|
1036
|
+
movq 16(%rbp), %rsi
|
|
1037
|
+
cmovzq %rsi, %rcx
|
|
1038
|
+
movq 24(%rbp), %rsi
|
|
1039
|
+
cmovzq %rsi, %rdx
|
|
1040
|
+
movq 32(%rbp), %rsi
|
|
1041
|
+
cmovzq %rsi, %r8
|
|
1042
|
+
movq 40(%rbp), %rsi
|
|
1043
|
+
cmovzq %rsi, %r9
|
|
1044
|
+
movq 48(%rbp), %rsi
|
|
1045
|
+
cmovzq %rsi, %r10
|
|
1046
|
+
movq 56(%rbp), %rsi
|
|
1047
|
+
cmovzq %rsi, %r11
|
|
1048
|
+
addq $128, %rbp
|
|
1049
|
+
|
|
1050
|
+
cmpq $8, bf
|
|
1051
|
+
movq (%rbp), %rsi
|
|
1052
|
+
cmovzq %rsi, %rax
|
|
1053
|
+
movq 8(%rbp), %rsi
|
|
1054
|
+
cmovzq %rsi, %rbx
|
|
1055
|
+
movq 16(%rbp), %rsi
|
|
1056
|
+
cmovzq %rsi, %rcx
|
|
1057
|
+
movq 24(%rbp), %rsi
|
|
1058
|
+
cmovzq %rsi, %rdx
|
|
1059
|
+
movq 32(%rbp), %rsi
|
|
1060
|
+
cmovzq %rsi, %r8
|
|
1061
|
+
movq 40(%rbp), %rsi
|
|
1062
|
+
cmovzq %rsi, %r9
|
|
1063
|
+
movq 48(%rbp), %rsi
|
|
1064
|
+
cmovzq %rsi, %r10
|
|
1065
|
+
movq 56(%rbp), %rsi
|
|
1066
|
+
cmovzq %rsi, %r11
|
|
1067
|
+
|
|
1068
|
+
movq %rax, TABENT+32(%rsp)
|
|
1069
|
+
movq %rbx, TABENT+40(%rsp)
|
|
1070
|
+
movq %rcx, TABENT+48(%rsp)
|
|
1071
|
+
movq %rdx, TABENT+56(%rsp)
|
|
1072
|
+
movq %r8, TABENT+64(%rsp)
|
|
1073
|
+
movq %r9, TABENT+72(%rsp)
|
|
1074
|
+
movq %r10, TABENT+80(%rsp)
|
|
1075
|
+
movq %r11, TABENT+88(%rsp)
|
|
1076
|
+
|
|
1077
|
+
// ...followed by the X and W fields
|
|
1078
|
+
|
|
1079
|
+
leaq TAB(%rsp), %rbp
|
|
1080
|
+
|
|
1081
|
+
xorl %eax, %eax
|
|
1082
|
+
xorl %ebx, %ebx
|
|
1083
|
+
xorl %ecx, %ecx
|
|
1084
|
+
xorl %edx, %edx
|
|
1085
|
+
xorl %r8d, %r8d
|
|
1086
|
+
xorl %r9d, %r9d
|
|
1087
|
+
xorl %r10d, %r10d
|
|
1088
|
+
xorl %r11d, %r11d
|
|
1089
|
+
|
|
1090
|
+
cmpq $1, bf
|
|
1091
|
+
movq (%rbp), %rsi
|
|
1092
|
+
cmovzq %rsi, %rax
|
|
1093
|
+
movq 8(%rbp), %rsi
|
|
1094
|
+
cmovzq %rsi, %rbx
|
|
1095
|
+
movq 16(%rbp), %rsi
|
|
1096
|
+
cmovzq %rsi, %rcx
|
|
1097
|
+
movq 24(%rbp), %rsi
|
|
1098
|
+
cmovzq %rsi, %rdx
|
|
1099
|
+
movq 96(%rbp), %rsi
|
|
1100
|
+
cmovzq %rsi, %r8
|
|
1101
|
+
movq 104(%rbp), %rsi
|
|
1102
|
+
cmovzq %rsi, %r9
|
|
1103
|
+
movq 112(%rbp), %rsi
|
|
1104
|
+
cmovzq %rsi, %r10
|
|
1105
|
+
movq 120(%rbp), %rsi
|
|
1106
|
+
cmovzq %rsi, %r11
|
|
1107
|
+
addq $128, %rbp
|
|
1108
|
+
|
|
1109
|
+
cmpq $2, bf
|
|
1110
|
+
movq (%rbp), %rsi
|
|
1111
|
+
cmovzq %rsi, %rax
|
|
1112
|
+
movq 8(%rbp), %rsi
|
|
1113
|
+
cmovzq %rsi, %rbx
|
|
1114
|
+
movq 16(%rbp), %rsi
|
|
1115
|
+
cmovzq %rsi, %rcx
|
|
1116
|
+
movq 24(%rbp), %rsi
|
|
1117
|
+
cmovzq %rsi, %rdx
|
|
1118
|
+
movq 96(%rbp), %rsi
|
|
1119
|
+
cmovzq %rsi, %r8
|
|
1120
|
+
movq 104(%rbp), %rsi
|
|
1121
|
+
cmovzq %rsi, %r9
|
|
1122
|
+
movq 112(%rbp), %rsi
|
|
1123
|
+
cmovzq %rsi, %r10
|
|
1124
|
+
movq 120(%rbp), %rsi
|
|
1125
|
+
cmovzq %rsi, %r11
|
|
1126
|
+
addq $128, %rbp
|
|
1127
|
+
|
|
1128
|
+
cmpq $3, bf
|
|
1129
|
+
movq (%rbp), %rsi
|
|
1130
|
+
cmovzq %rsi, %rax
|
|
1131
|
+
movq 8(%rbp), %rsi
|
|
1132
|
+
cmovzq %rsi, %rbx
|
|
1133
|
+
movq 16(%rbp), %rsi
|
|
1134
|
+
cmovzq %rsi, %rcx
|
|
1135
|
+
movq 24(%rbp), %rsi
|
|
1136
|
+
cmovzq %rsi, %rdx
|
|
1137
|
+
movq 96(%rbp), %rsi
|
|
1138
|
+
cmovzq %rsi, %r8
|
|
1139
|
+
movq 104(%rbp), %rsi
|
|
1140
|
+
cmovzq %rsi, %r9
|
|
1141
|
+
movq 112(%rbp), %rsi
|
|
1142
|
+
cmovzq %rsi, %r10
|
|
1143
|
+
movq 120(%rbp), %rsi
|
|
1144
|
+
cmovzq %rsi, %r11
|
|
1145
|
+
addq $128, %rbp
|
|
1146
|
+
|
|
1147
|
+
cmpq $4, bf
|
|
1148
|
+
movq (%rbp), %rsi
|
|
1149
|
+
cmovzq %rsi, %rax
|
|
1150
|
+
movq 8(%rbp), %rsi
|
|
1151
|
+
cmovzq %rsi, %rbx
|
|
1152
|
+
movq 16(%rbp), %rsi
|
|
1153
|
+
cmovzq %rsi, %rcx
|
|
1154
|
+
movq 24(%rbp), %rsi
|
|
1155
|
+
cmovzq %rsi, %rdx
|
|
1156
|
+
movq 96(%rbp), %rsi
|
|
1157
|
+
cmovzq %rsi, %r8
|
|
1158
|
+
movq 104(%rbp), %rsi
|
|
1159
|
+
cmovzq %rsi, %r9
|
|
1160
|
+
movq 112(%rbp), %rsi
|
|
1161
|
+
cmovzq %rsi, %r10
|
|
1162
|
+
movq 120(%rbp), %rsi
|
|
1163
|
+
cmovzq %rsi, %r11
|
|
1164
|
+
addq $128, %rbp
|
|
1165
|
+
|
|
1166
|
+
cmpq $5, bf
|
|
1167
|
+
movq (%rbp), %rsi
|
|
1168
|
+
cmovzq %rsi, %rax
|
|
1169
|
+
movq 8(%rbp), %rsi
|
|
1170
|
+
cmovzq %rsi, %rbx
|
|
1171
|
+
movq 16(%rbp), %rsi
|
|
1172
|
+
cmovzq %rsi, %rcx
|
|
1173
|
+
movq 24(%rbp), %rsi
|
|
1174
|
+
cmovzq %rsi, %rdx
|
|
1175
|
+
movq 96(%rbp), %rsi
|
|
1176
|
+
cmovzq %rsi, %r8
|
|
1177
|
+
movq 104(%rbp), %rsi
|
|
1178
|
+
cmovzq %rsi, %r9
|
|
1179
|
+
movq 112(%rbp), %rsi
|
|
1180
|
+
cmovzq %rsi, %r10
|
|
1181
|
+
movq 120(%rbp), %rsi
|
|
1182
|
+
cmovzq %rsi, %r11
|
|
1183
|
+
addq $128, %rbp
|
|
1184
|
+
|
|
1185
|
+
cmpq $6, bf
|
|
1186
|
+
movq (%rbp), %rsi
|
|
1187
|
+
cmovzq %rsi, %rax
|
|
1188
|
+
movq 8(%rbp), %rsi
|
|
1189
|
+
cmovzq %rsi, %rbx
|
|
1190
|
+
movq 16(%rbp), %rsi
|
|
1191
|
+
cmovzq %rsi, %rcx
|
|
1192
|
+
movq 24(%rbp), %rsi
|
|
1193
|
+
cmovzq %rsi, %rdx
|
|
1194
|
+
movq 96(%rbp), %rsi
|
|
1195
|
+
cmovzq %rsi, %r8
|
|
1196
|
+
movq 104(%rbp), %rsi
|
|
1197
|
+
cmovzq %rsi, %r9
|
|
1198
|
+
movq 112(%rbp), %rsi
|
|
1199
|
+
cmovzq %rsi, %r10
|
|
1200
|
+
movq 120(%rbp), %rsi
|
|
1201
|
+
cmovzq %rsi, %r11
|
|
1202
|
+
addq $128, %rbp
|
|
1203
|
+
|
|
1204
|
+
cmpq $7, bf
|
|
1205
|
+
movq (%rbp), %rsi
|
|
1206
|
+
cmovzq %rsi, %rax
|
|
1207
|
+
movq 8(%rbp), %rsi
|
|
1208
|
+
cmovzq %rsi, %rbx
|
|
1209
|
+
movq 16(%rbp), %rsi
|
|
1210
|
+
cmovzq %rsi, %rcx
|
|
1211
|
+
movq 24(%rbp), %rsi
|
|
1212
|
+
cmovzq %rsi, %rdx
|
|
1213
|
+
movq 96(%rbp), %rsi
|
|
1214
|
+
cmovzq %rsi, %r8
|
|
1215
|
+
movq 104(%rbp), %rsi
|
|
1216
|
+
cmovzq %rsi, %r9
|
|
1217
|
+
movq 112(%rbp), %rsi
|
|
1218
|
+
cmovzq %rsi, %r10
|
|
1219
|
+
movq 120(%rbp), %rsi
|
|
1220
|
+
cmovzq %rsi, %r11
|
|
1221
|
+
addq $128, %rbp
|
|
1222
|
+
|
|
1223
|
+
cmpq $8, bf
|
|
1224
|
+
movq (%rbp), %rsi
|
|
1225
|
+
cmovzq %rsi, %rax
|
|
1226
|
+
movq 8(%rbp), %rsi
|
|
1227
|
+
cmovzq %rsi, %rbx
|
|
1228
|
+
movq 16(%rbp), %rsi
|
|
1229
|
+
cmovzq %rsi, %rcx
|
|
1230
|
+
movq 24(%rbp), %rsi
|
|
1231
|
+
cmovzq %rsi, %rdx
|
|
1232
|
+
movq 96(%rbp), %rsi
|
|
1233
|
+
cmovzq %rsi, %r8
|
|
1234
|
+
movq 104(%rbp), %rsi
|
|
1235
|
+
cmovzq %rsi, %r9
|
|
1236
|
+
movq 112(%rbp), %rsi
|
|
1237
|
+
cmovzq %rsi, %r10
|
|
1238
|
+
movq 120(%rbp), %rsi
|
|
1239
|
+
cmovzq %rsi, %r11
|
|
1240
|
+
|
|
1241
|
+
movq %rax, TABENT(%rsp)
|
|
1242
|
+
movq %rbx, TABENT+8(%rsp)
|
|
1243
|
+
movq %rcx, TABENT+16(%rsp)
|
|
1244
|
+
movq %rdx, TABENT+24(%rsp)
|
|
1245
|
+
movq %r8, TABENT+96(%rsp)
|
|
1246
|
+
movq %r9, TABENT+104(%rsp)
|
|
1247
|
+
movq %r10, TABENT+112(%rsp)
|
|
1248
|
+
movq %r11, TABENT+120(%rsp)
|
|
1249
|
+
|
|
1250
|
+
// Add those elements to initialize the accumulator for bit position 252
|
|
1251
|
+
|
|
1252
|
+
leaq ACC(%rsp), %rdi
|
|
1253
|
+
leaq TABENT(%rsp), %rsi
|
|
1254
|
+
leaq BTABENT(%rsp), %rbp
|
|
1255
|
+
callq edwards25519_scalarmuldouble_pepadd
|
|
1256
|
+
|
|
1257
|
+
// Main loop with acc = [scalar/2^i] * point + [bscalar/2^i] * basepoint
|
|
1258
|
+
// Start with i = 252 for bits 248..251 and go down four at a time to 3..0
|
|
1259
|
+
|
|
1260
|
+
edwards25519_scalarmuldouble_loop:
|
|
1261
|
+
|
|
1262
|
+
movq i, %rax
|
|
1263
|
+
subq $4, %rax
|
|
1264
|
+
movq %rax, i
|
|
1265
|
+
|
|
1266
|
+
// Double to acc' = 2 * acc
|
|
1267
|
+
|
|
1268
|
+
leaq ACC(%rsp), %rdi
|
|
1269
|
+
leaq ACC(%rsp), %rsi
|
|
1270
|
+
callq edwards25519_scalarmuldouble_pdouble
|
|
1271
|
+
|
|
1272
|
+
// Get btable entry, first getting the adjusted bitfield...
|
|
1273
|
+
|
|
1274
|
+
movq i, %rax
|
|
1275
|
+
movq %rax, %rcx
|
|
1276
|
+
shrq $6, %rax
|
|
1277
|
+
movq 32(%rsp,%rax,8), %rax
|
|
1278
|
+
shrq %cl, %rax
|
|
1279
|
+
andq $15, %rax
|
|
1280
|
+
|
|
1281
|
+
subq $8, %rax
|
|
1282
|
+
sbbq %rcx, %rcx
|
|
1283
|
+
xorq %rcx, %rax
|
|
1284
|
+
subq %rcx, %rax
|
|
1285
|
+
movq %rcx, cf
|
|
1286
|
+
movq %rax, bf
|
|
1287
|
+
|
|
1288
|
+
// ... then doing constant-time lookup with the appropriate index...
|
|
1289
|
+
|
|
1290
|
+
movl $1, %eax
|
|
1291
|
+
xorl %ebx, %ebx
|
|
1292
|
+
xorl %ecx, %ecx
|
|
1293
|
+
xorl %edx, %edx
|
|
1294
|
+
movl $1, %r8d
|
|
1295
|
+
xorl %r9d, %r9d
|
|
1296
|
+
xorl %r10d, %r10d
|
|
1297
|
+
xorl %r11d, %r11d
|
|
1298
|
+
xorl %r12d, %r12d
|
|
1299
|
+
xorl %r13d, %r13d
|
|
1300
|
+
xorl %r14d, %r14d
|
|
1301
|
+
xorl %r15d, %r15d
|
|
1302
|
+
|
|
1303
|
+
leaq edwards25519_scalarmuldouble_table(%rip), %rbp
|
|
1304
|
+
|
|
1305
|
+
cmpq $1, bf
|
|
1306
|
+
movq (%rbp), %rsi
|
|
1307
|
+
cmovzq %rsi, %rax
|
|
1308
|
+
movq 8(%rbp), %rsi
|
|
1309
|
+
cmovzq %rsi, %rbx
|
|
1310
|
+
movq 16(%rbp), %rsi
|
|
1311
|
+
cmovzq %rsi, %rcx
|
|
1312
|
+
movq 24(%rbp), %rsi
|
|
1313
|
+
cmovzq %rsi, %rdx
|
|
1314
|
+
movq 32(%rbp), %rsi
|
|
1315
|
+
cmovzq %rsi, %r8
|
|
1316
|
+
movq 40(%rbp), %rsi
|
|
1317
|
+
cmovzq %rsi, %r9
|
|
1318
|
+
movq 48(%rbp), %rsi
|
|
1319
|
+
cmovzq %rsi, %r10
|
|
1320
|
+
movq 56(%rbp), %rsi
|
|
1321
|
+
cmovzq %rsi, %r11
|
|
1322
|
+
movq 64(%rbp), %rsi
|
|
1323
|
+
cmovzq %rsi, %r12
|
|
1324
|
+
movq 72(%rbp), %rsi
|
|
1325
|
+
cmovzq %rsi, %r13
|
|
1326
|
+
movq 80(%rbp), %rsi
|
|
1327
|
+
cmovzq %rsi, %r14
|
|
1328
|
+
movq 88(%rbp), %rsi
|
|
1329
|
+
cmovzq %rsi, %r15
|
|
1330
|
+
addq $96, %rbp
|
|
1331
|
+
|
|
1332
|
+
cmpq $2, bf
|
|
1333
|
+
movq (%rbp), %rsi
|
|
1334
|
+
cmovzq %rsi, %rax
|
|
1335
|
+
movq 8(%rbp), %rsi
|
|
1336
|
+
cmovzq %rsi, %rbx
|
|
1337
|
+
movq 16(%rbp), %rsi
|
|
1338
|
+
cmovzq %rsi, %rcx
|
|
1339
|
+
movq 24(%rbp), %rsi
|
|
1340
|
+
cmovzq %rsi, %rdx
|
|
1341
|
+
movq 32(%rbp), %rsi
|
|
1342
|
+
cmovzq %rsi, %r8
|
|
1343
|
+
movq 40(%rbp), %rsi
|
|
1344
|
+
cmovzq %rsi, %r9
|
|
1345
|
+
movq 48(%rbp), %rsi
|
|
1346
|
+
cmovzq %rsi, %r10
|
|
1347
|
+
movq 56(%rbp), %rsi
|
|
1348
|
+
cmovzq %rsi, %r11
|
|
1349
|
+
movq 64(%rbp), %rsi
|
|
1350
|
+
cmovzq %rsi, %r12
|
|
1351
|
+
movq 72(%rbp), %rsi
|
|
1352
|
+
cmovzq %rsi, %r13
|
|
1353
|
+
movq 80(%rbp), %rsi
|
|
1354
|
+
cmovzq %rsi, %r14
|
|
1355
|
+
movq 88(%rbp), %rsi
|
|
1356
|
+
cmovzq %rsi, %r15
|
|
1357
|
+
addq $96, %rbp
|
|
1358
|
+
|
|
1359
|
+
cmpq $3, bf
|
|
1360
|
+
movq (%rbp), %rsi
|
|
1361
|
+
cmovzq %rsi, %rax
|
|
1362
|
+
movq 8(%rbp), %rsi
|
|
1363
|
+
cmovzq %rsi, %rbx
|
|
1364
|
+
movq 16(%rbp), %rsi
|
|
1365
|
+
cmovzq %rsi, %rcx
|
|
1366
|
+
movq 24(%rbp), %rsi
|
|
1367
|
+
cmovzq %rsi, %rdx
|
|
1368
|
+
movq 32(%rbp), %rsi
|
|
1369
|
+
cmovzq %rsi, %r8
|
|
1370
|
+
movq 40(%rbp), %rsi
|
|
1371
|
+
cmovzq %rsi, %r9
|
|
1372
|
+
movq 48(%rbp), %rsi
|
|
1373
|
+
cmovzq %rsi, %r10
|
|
1374
|
+
movq 56(%rbp), %rsi
|
|
1375
|
+
cmovzq %rsi, %r11
|
|
1376
|
+
movq 64(%rbp), %rsi
|
|
1377
|
+
cmovzq %rsi, %r12
|
|
1378
|
+
movq 72(%rbp), %rsi
|
|
1379
|
+
cmovzq %rsi, %r13
|
|
1380
|
+
movq 80(%rbp), %rsi
|
|
1381
|
+
cmovzq %rsi, %r14
|
|
1382
|
+
movq 88(%rbp), %rsi
|
|
1383
|
+
cmovzq %rsi, %r15
|
|
1384
|
+
addq $96, %rbp
|
|
1385
|
+
|
|
1386
|
+
cmpq $4, bf
|
|
1387
|
+
movq (%rbp), %rsi
|
|
1388
|
+
cmovzq %rsi, %rax
|
|
1389
|
+
movq 8(%rbp), %rsi
|
|
1390
|
+
cmovzq %rsi, %rbx
|
|
1391
|
+
movq 16(%rbp), %rsi
|
|
1392
|
+
cmovzq %rsi, %rcx
|
|
1393
|
+
movq 24(%rbp), %rsi
|
|
1394
|
+
cmovzq %rsi, %rdx
|
|
1395
|
+
movq 32(%rbp), %rsi
|
|
1396
|
+
cmovzq %rsi, %r8
|
|
1397
|
+
movq 40(%rbp), %rsi
|
|
1398
|
+
cmovzq %rsi, %r9
|
|
1399
|
+
movq 48(%rbp), %rsi
|
|
1400
|
+
cmovzq %rsi, %r10
|
|
1401
|
+
movq 56(%rbp), %rsi
|
|
1402
|
+
cmovzq %rsi, %r11
|
|
1403
|
+
movq 64(%rbp), %rsi
|
|
1404
|
+
cmovzq %rsi, %r12
|
|
1405
|
+
movq 72(%rbp), %rsi
|
|
1406
|
+
cmovzq %rsi, %r13
|
|
1407
|
+
movq 80(%rbp), %rsi
|
|
1408
|
+
cmovzq %rsi, %r14
|
|
1409
|
+
movq 88(%rbp), %rsi
|
|
1410
|
+
cmovzq %rsi, %r15
|
|
1411
|
+
addq $96, %rbp
|
|
1412
|
+
|
|
1413
|
+
cmpq $5, bf
|
|
1414
|
+
movq (%rbp), %rsi
|
|
1415
|
+
cmovzq %rsi, %rax
|
|
1416
|
+
movq 8(%rbp), %rsi
|
|
1417
|
+
cmovzq %rsi, %rbx
|
|
1418
|
+
movq 16(%rbp), %rsi
|
|
1419
|
+
cmovzq %rsi, %rcx
|
|
1420
|
+
movq 24(%rbp), %rsi
|
|
1421
|
+
cmovzq %rsi, %rdx
|
|
1422
|
+
movq 32(%rbp), %rsi
|
|
1423
|
+
cmovzq %rsi, %r8
|
|
1424
|
+
movq 40(%rbp), %rsi
|
|
1425
|
+
cmovzq %rsi, %r9
|
|
1426
|
+
movq 48(%rbp), %rsi
|
|
1427
|
+
cmovzq %rsi, %r10
|
|
1428
|
+
movq 56(%rbp), %rsi
|
|
1429
|
+
cmovzq %rsi, %r11
|
|
1430
|
+
movq 64(%rbp), %rsi
|
|
1431
|
+
cmovzq %rsi, %r12
|
|
1432
|
+
movq 72(%rbp), %rsi
|
|
1433
|
+
cmovzq %rsi, %r13
|
|
1434
|
+
movq 80(%rbp), %rsi
|
|
1435
|
+
cmovzq %rsi, %r14
|
|
1436
|
+
movq 88(%rbp), %rsi
|
|
1437
|
+
cmovzq %rsi, %r15
|
|
1438
|
+
addq $96, %rbp
|
|
1439
|
+
|
|
1440
|
+
cmpq $6, bf
|
|
1441
|
+
movq (%rbp), %rsi
|
|
1442
|
+
cmovzq %rsi, %rax
|
|
1443
|
+
movq 8(%rbp), %rsi
|
|
1444
|
+
cmovzq %rsi, %rbx
|
|
1445
|
+
movq 16(%rbp), %rsi
|
|
1446
|
+
cmovzq %rsi, %rcx
|
|
1447
|
+
movq 24(%rbp), %rsi
|
|
1448
|
+
cmovzq %rsi, %rdx
|
|
1449
|
+
movq 32(%rbp), %rsi
|
|
1450
|
+
cmovzq %rsi, %r8
|
|
1451
|
+
movq 40(%rbp), %rsi
|
|
1452
|
+
cmovzq %rsi, %r9
|
|
1453
|
+
movq 48(%rbp), %rsi
|
|
1454
|
+
cmovzq %rsi, %r10
|
|
1455
|
+
movq 56(%rbp), %rsi
|
|
1456
|
+
cmovzq %rsi, %r11
|
|
1457
|
+
movq 64(%rbp), %rsi
|
|
1458
|
+
cmovzq %rsi, %r12
|
|
1459
|
+
movq 72(%rbp), %rsi
|
|
1460
|
+
cmovzq %rsi, %r13
|
|
1461
|
+
movq 80(%rbp), %rsi
|
|
1462
|
+
cmovzq %rsi, %r14
|
|
1463
|
+
movq 88(%rbp), %rsi
|
|
1464
|
+
cmovzq %rsi, %r15
|
|
1465
|
+
addq $96, %rbp
|
|
1466
|
+
|
|
1467
|
+
cmpq $7, bf
|
|
1468
|
+
movq (%rbp), %rsi
|
|
1469
|
+
cmovzq %rsi, %rax
|
|
1470
|
+
movq 8(%rbp), %rsi
|
|
1471
|
+
cmovzq %rsi, %rbx
|
|
1472
|
+
movq 16(%rbp), %rsi
|
|
1473
|
+
cmovzq %rsi, %rcx
|
|
1474
|
+
movq 24(%rbp), %rsi
|
|
1475
|
+
cmovzq %rsi, %rdx
|
|
1476
|
+
movq 32(%rbp), %rsi
|
|
1477
|
+
cmovzq %rsi, %r8
|
|
1478
|
+
movq 40(%rbp), %rsi
|
|
1479
|
+
cmovzq %rsi, %r9
|
|
1480
|
+
movq 48(%rbp), %rsi
|
|
1481
|
+
cmovzq %rsi, %r10
|
|
1482
|
+
movq 56(%rbp), %rsi
|
|
1483
|
+
cmovzq %rsi, %r11
|
|
1484
|
+
movq 64(%rbp), %rsi
|
|
1485
|
+
cmovzq %rsi, %r12
|
|
1486
|
+
movq 72(%rbp), %rsi
|
|
1487
|
+
cmovzq %rsi, %r13
|
|
1488
|
+
movq 80(%rbp), %rsi
|
|
1489
|
+
cmovzq %rsi, %r14
|
|
1490
|
+
movq 88(%rbp), %rsi
|
|
1491
|
+
cmovzq %rsi, %r15
|
|
1492
|
+
addq $96, %rbp
|
|
1493
|
+
|
|
1494
|
+
cmpq $8, bf
|
|
1495
|
+
movq (%rbp), %rsi
|
|
1496
|
+
cmovzq %rsi, %rax
|
|
1497
|
+
movq 8(%rbp), %rsi
|
|
1498
|
+
cmovzq %rsi, %rbx
|
|
1499
|
+
movq 16(%rbp), %rsi
|
|
1500
|
+
cmovzq %rsi, %rcx
|
|
1501
|
+
movq 24(%rbp), %rsi
|
|
1502
|
+
cmovzq %rsi, %rdx
|
|
1503
|
+
movq 32(%rbp), %rsi
|
|
1504
|
+
cmovzq %rsi, %r8
|
|
1505
|
+
movq 40(%rbp), %rsi
|
|
1506
|
+
cmovzq %rsi, %r9
|
|
1507
|
+
movq 48(%rbp), %rsi
|
|
1508
|
+
cmovzq %rsi, %r10
|
|
1509
|
+
movq 56(%rbp), %rsi
|
|
1510
|
+
cmovzq %rsi, %r11
|
|
1511
|
+
movq 64(%rbp), %rsi
|
|
1512
|
+
cmovzq %rsi, %r12
|
|
1513
|
+
movq 72(%rbp), %rsi
|
|
1514
|
+
cmovzq %rsi, %r13
|
|
1515
|
+
movq 80(%rbp), %rsi
|
|
1516
|
+
cmovzq %rsi, %r14
|
|
1517
|
+
movq 88(%rbp), %rsi
|
|
1518
|
+
cmovzq %rsi, %r15
|
|
1519
|
+
|
|
1520
|
+
// ... then optionally negating before storing. The table entry
|
|
1521
|
+
// is in precomputed form and we currently have
|
|
1522
|
+
//
|
|
1523
|
+
// [%rdx;%rcx;%rbx;%rax] = y - x
|
|
1524
|
+
// [%r11;%r10;%r9;%r8] = x + y
|
|
1525
|
+
// [%r15;%r14;%r13;%r12] = 2 * d * x * y
|
|
1526
|
+
//
|
|
1527
|
+
// Negation for Edwards curves is -(x,y) = (-x,y), which in this modified
|
|
1528
|
+
// form amounts to swapping the first two fields and negating the third.
|
|
1529
|
+
// The negation does not always fully reduce even mod 2^256-38 in the zero
|
|
1530
|
+
// case, instead giving -0 = 2^256-38. But that is fine since the result is
|
|
1531
|
+
// always fed to a multiplication inside the "pepadd" function below that
|
|
1532
|
+
// handles any 256-bit input.
|
|
1533
|
+
|
|
1534
|
+
movq cf, %rdi
|
|
1535
|
+
testq %rdi, %rdi
|
|
1536
|
+
|
|
1537
|
+
movq %rax, %rsi
|
|
1538
|
+
cmovnzq %r8, %rsi
|
|
1539
|
+
cmovnzq %rax, %r8
|
|
1540
|
+
movq %rsi, BTABENT(%rsp)
|
|
1541
|
+
movq %r8, BTABENT+32(%rsp)
|
|
1542
|
+
|
|
1543
|
+
movq %rbx, %rsi
|
|
1544
|
+
cmovnzq %r9, %rsi
|
|
1545
|
+
cmovnzq %rbx, %r9
|
|
1546
|
+
movq %rsi, BTABENT+8(%rsp)
|
|
1547
|
+
movq %r9, BTABENT+40(%rsp)
|
|
1548
|
+
|
|
1549
|
+
movq %rcx, %rsi
|
|
1550
|
+
cmovnzq %r10, %rsi
|
|
1551
|
+
cmovnzq %rcx, %r10
|
|
1552
|
+
movq %rsi, BTABENT+16(%rsp)
|
|
1553
|
+
movq %r10, BTABENT+48(%rsp)
|
|
1554
|
+
|
|
1555
|
+
movq %rdx, %rsi
|
|
1556
|
+
cmovnzq %r11, %rsi
|
|
1557
|
+
cmovnzq %rdx, %r11
|
|
1558
|
+
movq %rsi, BTABENT+24(%rsp)
|
|
1559
|
+
movq %r11, BTABENT+56(%rsp)
|
|
1560
|
+
|
|
1561
|
+
xorq %rdi, %r12
|
|
1562
|
+
xorq %rdi, %r13
|
|
1563
|
+
xorq %rdi, %r14
|
|
1564
|
+
xorq %rdi, %r15
|
|
1565
|
+
andq $37, %rdi
|
|
1566
|
+
subq %rdi, %r12
|
|
1567
|
+
sbbq $0, %r13
|
|
1568
|
+
sbbq $0, %r14
|
|
1569
|
+
sbbq $0, %r15
|
|
1570
|
+
movq %r12, BTABENT+64(%rsp)
|
|
1571
|
+
movq %r13, BTABENT+72(%rsp)
|
|
1572
|
+
movq %r14, BTABENT+80(%rsp)
|
|
1573
|
+
movq %r15, BTABENT+88(%rsp)
|
|
1574
|
+
|
|
1575
|
+
// Get table entry, first getting the adjusted bitfield...
|
|
1576
|
+
|
|
1577
|
+
movq i, %rax
|
|
1578
|
+
movq %rax, %rcx
|
|
1579
|
+
shrq $6, %rax
|
|
1580
|
+
movq (%rsp,%rax,8), %rax
|
|
1581
|
+
shrq %cl, %rax
|
|
1582
|
+
andq $15, %rax
|
|
1583
|
+
|
|
1584
|
+
subq $8, %rax
|
|
1585
|
+
sbbq %rcx, %rcx
|
|
1586
|
+
xorq %rcx, %rax
|
|
1587
|
+
subq %rcx, %rax
|
|
1588
|
+
movq %rcx, cf
|
|
1589
|
+
movq %rax, bf
|
|
1590
|
+
|
|
1591
|
+
// ...and constant-time indexing based on that index
|
|
1592
|
+
// Do the Y and Z fields first, to save on registers
|
|
1593
|
+
// and store them back (they don't need any modification)
|
|
1594
|
+
|
|
1595
|
+
movl $1, %eax
|
|
1596
|
+
xorl %ebx, %ebx
|
|
1597
|
+
xorl %ecx, %ecx
|
|
1598
|
+
xorl %edx, %edx
|
|
1599
|
+
movl $1, %r8d
|
|
1600
|
+
xorl %r9d, %r9d
|
|
1601
|
+
xorl %r10d, %r10d
|
|
1602
|
+
xorl %r11d, %r11d
|
|
1603
|
+
|
|
1604
|
+
leaq TAB+32(%rsp), %rbp
|
|
1605
|
+
|
|
1606
|
+
cmpq $1, bf
|
|
1607
|
+
movq (%rbp), %rsi
|
|
1608
|
+
cmovzq %rsi, %rax
|
|
1609
|
+
movq 8(%rbp), %rsi
|
|
1610
|
+
cmovzq %rsi, %rbx
|
|
1611
|
+
movq 16(%rbp), %rsi
|
|
1612
|
+
cmovzq %rsi, %rcx
|
|
1613
|
+
movq 24(%rbp), %rsi
|
|
1614
|
+
cmovzq %rsi, %rdx
|
|
1615
|
+
movq 32(%rbp), %rsi
|
|
1616
|
+
cmovzq %rsi, %r8
|
|
1617
|
+
movq 40(%rbp), %rsi
|
|
1618
|
+
cmovzq %rsi, %r9
|
|
1619
|
+
movq 48(%rbp), %rsi
|
|
1620
|
+
cmovzq %rsi, %r10
|
|
1621
|
+
movq 56(%rbp), %rsi
|
|
1622
|
+
cmovzq %rsi, %r11
|
|
1623
|
+
addq $128, %rbp
|
|
1624
|
+
|
|
1625
|
+
cmpq $2, bf
|
|
1626
|
+
movq (%rbp), %rsi
|
|
1627
|
+
cmovzq %rsi, %rax
|
|
1628
|
+
movq 8(%rbp), %rsi
|
|
1629
|
+
cmovzq %rsi, %rbx
|
|
1630
|
+
movq 16(%rbp), %rsi
|
|
1631
|
+
cmovzq %rsi, %rcx
|
|
1632
|
+
movq 24(%rbp), %rsi
|
|
1633
|
+
cmovzq %rsi, %rdx
|
|
1634
|
+
movq 32(%rbp), %rsi
|
|
1635
|
+
cmovzq %rsi, %r8
|
|
1636
|
+
movq 40(%rbp), %rsi
|
|
1637
|
+
cmovzq %rsi, %r9
|
|
1638
|
+
movq 48(%rbp), %rsi
|
|
1639
|
+
cmovzq %rsi, %r10
|
|
1640
|
+
movq 56(%rbp), %rsi
|
|
1641
|
+
cmovzq %rsi, %r11
|
|
1642
|
+
addq $128, %rbp
|
|
1643
|
+
|
|
1644
|
+
cmpq $3, bf
|
|
1645
|
+
movq (%rbp), %rsi
|
|
1646
|
+
cmovzq %rsi, %rax
|
|
1647
|
+
movq 8(%rbp), %rsi
|
|
1648
|
+
cmovzq %rsi, %rbx
|
|
1649
|
+
movq 16(%rbp), %rsi
|
|
1650
|
+
cmovzq %rsi, %rcx
|
|
1651
|
+
movq 24(%rbp), %rsi
|
|
1652
|
+
cmovzq %rsi, %rdx
|
|
1653
|
+
movq 32(%rbp), %rsi
|
|
1654
|
+
cmovzq %rsi, %r8
|
|
1655
|
+
movq 40(%rbp), %rsi
|
|
1656
|
+
cmovzq %rsi, %r9
|
|
1657
|
+
movq 48(%rbp), %rsi
|
|
1658
|
+
cmovzq %rsi, %r10
|
|
1659
|
+
movq 56(%rbp), %rsi
|
|
1660
|
+
cmovzq %rsi, %r11
|
|
1661
|
+
addq $128, %rbp
|
|
1662
|
+
|
|
1663
|
+
cmpq $4, bf
|
|
1664
|
+
movq (%rbp), %rsi
|
|
1665
|
+
cmovzq %rsi, %rax
|
|
1666
|
+
movq 8(%rbp), %rsi
|
|
1667
|
+
cmovzq %rsi, %rbx
|
|
1668
|
+
movq 16(%rbp), %rsi
|
|
1669
|
+
cmovzq %rsi, %rcx
|
|
1670
|
+
movq 24(%rbp), %rsi
|
|
1671
|
+
cmovzq %rsi, %rdx
|
|
1672
|
+
movq 32(%rbp), %rsi
|
|
1673
|
+
cmovzq %rsi, %r8
|
|
1674
|
+
movq 40(%rbp), %rsi
|
|
1675
|
+
cmovzq %rsi, %r9
|
|
1676
|
+
movq 48(%rbp), %rsi
|
|
1677
|
+
cmovzq %rsi, %r10
|
|
1678
|
+
movq 56(%rbp), %rsi
|
|
1679
|
+
cmovzq %rsi, %r11
|
|
1680
|
+
addq $128, %rbp
|
|
1681
|
+
|
|
1682
|
+
cmpq $5, bf
|
|
1683
|
+
movq (%rbp), %rsi
|
|
1684
|
+
cmovzq %rsi, %rax
|
|
1685
|
+
movq 8(%rbp), %rsi
|
|
1686
|
+
cmovzq %rsi, %rbx
|
|
1687
|
+
movq 16(%rbp), %rsi
|
|
1688
|
+
cmovzq %rsi, %rcx
|
|
1689
|
+
movq 24(%rbp), %rsi
|
|
1690
|
+
cmovzq %rsi, %rdx
|
|
1691
|
+
movq 32(%rbp), %rsi
|
|
1692
|
+
cmovzq %rsi, %r8
|
|
1693
|
+
movq 40(%rbp), %rsi
|
|
1694
|
+
cmovzq %rsi, %r9
|
|
1695
|
+
movq 48(%rbp), %rsi
|
|
1696
|
+
cmovzq %rsi, %r10
|
|
1697
|
+
movq 56(%rbp), %rsi
|
|
1698
|
+
cmovzq %rsi, %r11
|
|
1699
|
+
addq $128, %rbp
|
|
1700
|
+
|
|
1701
|
+
cmpq $6, bf
|
|
1702
|
+
movq (%rbp), %rsi
|
|
1703
|
+
cmovzq %rsi, %rax
|
|
1704
|
+
movq 8(%rbp), %rsi
|
|
1705
|
+
cmovzq %rsi, %rbx
|
|
1706
|
+
movq 16(%rbp), %rsi
|
|
1707
|
+
cmovzq %rsi, %rcx
|
|
1708
|
+
movq 24(%rbp), %rsi
|
|
1709
|
+
cmovzq %rsi, %rdx
|
|
1710
|
+
movq 32(%rbp), %rsi
|
|
1711
|
+
cmovzq %rsi, %r8
|
|
1712
|
+
movq 40(%rbp), %rsi
|
|
1713
|
+
cmovzq %rsi, %r9
|
|
1714
|
+
movq 48(%rbp), %rsi
|
|
1715
|
+
cmovzq %rsi, %r10
|
|
1716
|
+
movq 56(%rbp), %rsi
|
|
1717
|
+
cmovzq %rsi, %r11
|
|
1718
|
+
addq $128, %rbp
|
|
1719
|
+
|
|
1720
|
+
cmpq $7, bf
|
|
1721
|
+
movq (%rbp), %rsi
|
|
1722
|
+
cmovzq %rsi, %rax
|
|
1723
|
+
movq 8(%rbp), %rsi
|
|
1724
|
+
cmovzq %rsi, %rbx
|
|
1725
|
+
movq 16(%rbp), %rsi
|
|
1726
|
+
cmovzq %rsi, %rcx
|
|
1727
|
+
movq 24(%rbp), %rsi
|
|
1728
|
+
cmovzq %rsi, %rdx
|
|
1729
|
+
movq 32(%rbp), %rsi
|
|
1730
|
+
cmovzq %rsi, %r8
|
|
1731
|
+
movq 40(%rbp), %rsi
|
|
1732
|
+
cmovzq %rsi, %r9
|
|
1733
|
+
movq 48(%rbp), %rsi
|
|
1734
|
+
cmovzq %rsi, %r10
|
|
1735
|
+
movq 56(%rbp), %rsi
|
|
1736
|
+
cmovzq %rsi, %r11
|
|
1737
|
+
addq $128, %rbp
|
|
1738
|
+
|
|
1739
|
+
cmpq $8, bf
|
|
1740
|
+
movq (%rbp), %rsi
|
|
1741
|
+
cmovzq %rsi, %rax
|
|
1742
|
+
movq 8(%rbp), %rsi
|
|
1743
|
+
cmovzq %rsi, %rbx
|
|
1744
|
+
movq 16(%rbp), %rsi
|
|
1745
|
+
cmovzq %rsi, %rcx
|
|
1746
|
+
movq 24(%rbp), %rsi
|
|
1747
|
+
cmovzq %rsi, %rdx
|
|
1748
|
+
movq 32(%rbp), %rsi
|
|
1749
|
+
cmovzq %rsi, %r8
|
|
1750
|
+
movq 40(%rbp), %rsi
|
|
1751
|
+
cmovzq %rsi, %r9
|
|
1752
|
+
movq 48(%rbp), %rsi
|
|
1753
|
+
cmovzq %rsi, %r10
|
|
1754
|
+
movq 56(%rbp), %rsi
|
|
1755
|
+
cmovzq %rsi, %r11
|
|
1756
|
+
|
|
1757
|
+
movq %rax, TABENT+32(%rsp)
|
|
1758
|
+
movq %rbx, TABENT+40(%rsp)
|
|
1759
|
+
movq %rcx, TABENT+48(%rsp)
|
|
1760
|
+
movq %rdx, TABENT+56(%rsp)
|
|
1761
|
+
movq %r8, TABENT+64(%rsp)
|
|
1762
|
+
movq %r9, TABENT+72(%rsp)
|
|
1763
|
+
movq %r10, TABENT+80(%rsp)
|
|
1764
|
+
movq %r11, TABENT+88(%rsp)
|
|
1765
|
+
|
|
1766
|
+
// Now do the X and W fields...
|
|
1767
|
+
|
|
1768
|
+
leaq TAB(%rsp), %rbp
|
|
1769
|
+
|
|
1770
|
+
xorl %eax, %eax
|
|
1771
|
+
xorl %ebx, %ebx
|
|
1772
|
+
xorl %ecx, %ecx
|
|
1773
|
+
xorl %edx, %edx
|
|
1774
|
+
xorl %r8d, %r8d
|
|
1775
|
+
xorl %r9d, %r9d
|
|
1776
|
+
xorl %r10d, %r10d
|
|
1777
|
+
xorl %r11d, %r11d
|
|
1778
|
+
|
|
1779
|
+
cmpq $1, bf
|
|
1780
|
+
movq (%rbp), %rsi
|
|
1781
|
+
cmovzq %rsi, %rax
|
|
1782
|
+
movq 8(%rbp), %rsi
|
|
1783
|
+
cmovzq %rsi, %rbx
|
|
1784
|
+
movq 16(%rbp), %rsi
|
|
1785
|
+
cmovzq %rsi, %rcx
|
|
1786
|
+
movq 24(%rbp), %rsi
|
|
1787
|
+
cmovzq %rsi, %rdx
|
|
1788
|
+
movq 96(%rbp), %rsi
|
|
1789
|
+
cmovzq %rsi, %r8
|
|
1790
|
+
movq 104(%rbp), %rsi
|
|
1791
|
+
cmovzq %rsi, %r9
|
|
1792
|
+
movq 112(%rbp), %rsi
|
|
1793
|
+
cmovzq %rsi, %r10
|
|
1794
|
+
movq 120(%rbp), %rsi
|
|
1795
|
+
cmovzq %rsi, %r11
|
|
1796
|
+
addq $128, %rbp
|
|
1797
|
+
|
|
1798
|
+
cmpq $2, bf
|
|
1799
|
+
movq (%rbp), %rsi
|
|
1800
|
+
cmovzq %rsi, %rax
|
|
1801
|
+
movq 8(%rbp), %rsi
|
|
1802
|
+
cmovzq %rsi, %rbx
|
|
1803
|
+
movq 16(%rbp), %rsi
|
|
1804
|
+
cmovzq %rsi, %rcx
|
|
1805
|
+
movq 24(%rbp), %rsi
|
|
1806
|
+
cmovzq %rsi, %rdx
|
|
1807
|
+
movq 96(%rbp), %rsi
|
|
1808
|
+
cmovzq %rsi, %r8
|
|
1809
|
+
movq 104(%rbp), %rsi
|
|
1810
|
+
cmovzq %rsi, %r9
|
|
1811
|
+
movq 112(%rbp), %rsi
|
|
1812
|
+
cmovzq %rsi, %r10
|
|
1813
|
+
movq 120(%rbp), %rsi
|
|
1814
|
+
cmovzq %rsi, %r11
|
|
1815
|
+
addq $128, %rbp
|
|
1816
|
+
|
|
1817
|
+
cmpq $3, bf
|
|
1818
|
+
movq (%rbp), %rsi
|
|
1819
|
+
cmovzq %rsi, %rax
|
|
1820
|
+
movq 8(%rbp), %rsi
|
|
1821
|
+
cmovzq %rsi, %rbx
|
|
1822
|
+
movq 16(%rbp), %rsi
|
|
1823
|
+
cmovzq %rsi, %rcx
|
|
1824
|
+
movq 24(%rbp), %rsi
|
|
1825
|
+
cmovzq %rsi, %rdx
|
|
1826
|
+
movq 96(%rbp), %rsi
|
|
1827
|
+
cmovzq %rsi, %r8
|
|
1828
|
+
movq 104(%rbp), %rsi
|
|
1829
|
+
cmovzq %rsi, %r9
|
|
1830
|
+
movq 112(%rbp), %rsi
|
|
1831
|
+
cmovzq %rsi, %r10
|
|
1832
|
+
movq 120(%rbp), %rsi
|
|
1833
|
+
cmovzq %rsi, %r11
|
|
1834
|
+
addq $128, %rbp
|
|
1835
|
+
|
|
1836
|
+
cmpq $4, bf
|
|
1837
|
+
movq (%rbp), %rsi
|
|
1838
|
+
cmovzq %rsi, %rax
|
|
1839
|
+
movq 8(%rbp), %rsi
|
|
1840
|
+
cmovzq %rsi, %rbx
|
|
1841
|
+
movq 16(%rbp), %rsi
|
|
1842
|
+
cmovzq %rsi, %rcx
|
|
1843
|
+
movq 24(%rbp), %rsi
|
|
1844
|
+
cmovzq %rsi, %rdx
|
|
1845
|
+
movq 96(%rbp), %rsi
|
|
1846
|
+
cmovzq %rsi, %r8
|
|
1847
|
+
movq 104(%rbp), %rsi
|
|
1848
|
+
cmovzq %rsi, %r9
|
|
1849
|
+
movq 112(%rbp), %rsi
|
|
1850
|
+
cmovzq %rsi, %r10
|
|
1851
|
+
movq 120(%rbp), %rsi
|
|
1852
|
+
cmovzq %rsi, %r11
|
|
1853
|
+
addq $128, %rbp
|
|
1854
|
+
|
|
1855
|
+
cmpq $5, bf
|
|
1856
|
+
movq (%rbp), %rsi
|
|
1857
|
+
cmovzq %rsi, %rax
|
|
1858
|
+
movq 8(%rbp), %rsi
|
|
1859
|
+
cmovzq %rsi, %rbx
|
|
1860
|
+
movq 16(%rbp), %rsi
|
|
1861
|
+
cmovzq %rsi, %rcx
|
|
1862
|
+
movq 24(%rbp), %rsi
|
|
1863
|
+
cmovzq %rsi, %rdx
|
|
1864
|
+
movq 96(%rbp), %rsi
|
|
1865
|
+
cmovzq %rsi, %r8
|
|
1866
|
+
movq 104(%rbp), %rsi
|
|
1867
|
+
cmovzq %rsi, %r9
|
|
1868
|
+
movq 112(%rbp), %rsi
|
|
1869
|
+
cmovzq %rsi, %r10
|
|
1870
|
+
movq 120(%rbp), %rsi
|
|
1871
|
+
cmovzq %rsi, %r11
|
|
1872
|
+
addq $128, %rbp
|
|
1873
|
+
|
|
1874
|
+
cmpq $6, bf
|
|
1875
|
+
movq (%rbp), %rsi
|
|
1876
|
+
cmovzq %rsi, %rax
|
|
1877
|
+
movq 8(%rbp), %rsi
|
|
1878
|
+
cmovzq %rsi, %rbx
|
|
1879
|
+
movq 16(%rbp), %rsi
|
|
1880
|
+
cmovzq %rsi, %rcx
|
|
1881
|
+
movq 24(%rbp), %rsi
|
|
1882
|
+
cmovzq %rsi, %rdx
|
|
1883
|
+
movq 96(%rbp), %rsi
|
|
1884
|
+
cmovzq %rsi, %r8
|
|
1885
|
+
movq 104(%rbp), %rsi
|
|
1886
|
+
cmovzq %rsi, %r9
|
|
1887
|
+
movq 112(%rbp), %rsi
|
|
1888
|
+
cmovzq %rsi, %r10
|
|
1889
|
+
movq 120(%rbp), %rsi
|
|
1890
|
+
cmovzq %rsi, %r11
|
|
1891
|
+
addq $128, %rbp
|
|
1892
|
+
|
|
1893
|
+
cmpq $7, bf
|
|
1894
|
+
movq (%rbp), %rsi
|
|
1895
|
+
cmovzq %rsi, %rax
|
|
1896
|
+
movq 8(%rbp), %rsi
|
|
1897
|
+
cmovzq %rsi, %rbx
|
|
1898
|
+
movq 16(%rbp), %rsi
|
|
1899
|
+
cmovzq %rsi, %rcx
|
|
1900
|
+
movq 24(%rbp), %rsi
|
|
1901
|
+
cmovzq %rsi, %rdx
|
|
1902
|
+
movq 96(%rbp), %rsi
|
|
1903
|
+
cmovzq %rsi, %r8
|
|
1904
|
+
movq 104(%rbp), %rsi
|
|
1905
|
+
cmovzq %rsi, %r9
|
|
1906
|
+
movq 112(%rbp), %rsi
|
|
1907
|
+
cmovzq %rsi, %r10
|
|
1908
|
+
movq 120(%rbp), %rsi
|
|
1909
|
+
cmovzq %rsi, %r11
|
|
1910
|
+
addq $128, %rbp
|
|
1911
|
+
|
|
1912
|
+
cmpq $8, bf
|
|
1913
|
+
movq (%rbp), %rsi
|
|
1914
|
+
cmovzq %rsi, %rax
|
|
1915
|
+
movq 8(%rbp), %rsi
|
|
1916
|
+
cmovzq %rsi, %rbx
|
|
1917
|
+
movq 16(%rbp), %rsi
|
|
1918
|
+
cmovzq %rsi, %rcx
|
|
1919
|
+
movq 24(%rbp), %rsi
|
|
1920
|
+
cmovzq %rsi, %rdx
|
|
1921
|
+
movq 96(%rbp), %rsi
|
|
1922
|
+
cmovzq %rsi, %r8
|
|
1923
|
+
movq 104(%rbp), %rsi
|
|
1924
|
+
cmovzq %rsi, %r9
|
|
1925
|
+
movq 112(%rbp), %rsi
|
|
1926
|
+
cmovzq %rsi, %r10
|
|
1927
|
+
movq 120(%rbp), %rsi
|
|
1928
|
+
cmovzq %rsi, %r11
|
|
1929
|
+
|
|
1930
|
+
// ... then optionally negate before storing the X and W fields. This
|
|
1931
|
+
// time the table entry is extended-projective, and is here:
|
|
1932
|
+
//
|
|
1933
|
+
// [%rdx;%rcx;%rbx;%rax] = X
|
|
1934
|
+
// [tabent+32] = Y
|
|
1935
|
+
// [tabent+64] = Z
|
|
1936
|
+
// [%r11;%r10;%r9;%r8] = W
|
|
1937
|
+
//
|
|
1938
|
+
// This time we just need to negate the X and the W fields.
|
|
1939
|
+
// The crude way negation is done can result in values of X or W
|
|
1940
|
+
// (when initially zero before negation) being exactly equal to
|
|
1941
|
+
// 2^256-38, but the "pepadd" function handles that correctly.
|
|
1942
|
+
|
|
1943
|
+
movq cf, %rdi
|
|
1944
|
+
|
|
1945
|
+
xorq %rdi, %rax
|
|
1946
|
+
xorq %rdi, %rbx
|
|
1947
|
+
xorq %rdi, %rcx
|
|
1948
|
+
xorq %rdi, %rdx
|
|
1949
|
+
|
|
1950
|
+
xorq %rdi, %r8
|
|
1951
|
+
xorq %rdi, %r9
|
|
1952
|
+
xorq %rdi, %r10
|
|
1953
|
+
xorq %rdi, %r11
|
|
1954
|
+
|
|
1955
|
+
andq $37, %rdi
|
|
1956
|
+
|
|
1957
|
+
subq %rdi, %rax
|
|
1958
|
+
sbbq $0, %rbx
|
|
1959
|
+
sbbq $0, %rcx
|
|
1960
|
+
sbbq $0, %rdx
|
|
1961
|
+
|
|
1962
|
+
movq %rax, TABENT(%rsp)
|
|
1963
|
+
movq %rbx, TABENT+8(%rsp)
|
|
1964
|
+
movq %rcx, TABENT+16(%rsp)
|
|
1965
|
+
movq %rdx, TABENT+24(%rsp)
|
|
1966
|
+
|
|
1967
|
+
subq %rdi, %r8
|
|
1968
|
+
sbbq $0, %r9
|
|
1969
|
+
sbbq $0, %r10
|
|
1970
|
+
sbbq $0, %r11
|
|
1971
|
+
|
|
1972
|
+
movq %r8, TABENT+96(%rsp)
|
|
1973
|
+
movq %r9, TABENT+104(%rsp)
|
|
1974
|
+
movq %r10, TABENT+112(%rsp)
|
|
1975
|
+
movq %r11, TABENT+120(%rsp)
|
|
1976
|
+
|
|
1977
|
+
// Double to acc' = 4 * acc
|
|
1978
|
+
|
|
1979
|
+
leaq ACC(%rsp), %rdi
|
|
1980
|
+
leaq ACC(%rsp), %rsi
|
|
1981
|
+
callq edwards25519_scalarmuldouble_pdouble
|
|
1982
|
+
|
|
1983
|
+
// Add tabent := tabent + btabent
|
|
1984
|
+
|
|
1985
|
+
leaq TABENT(%rsp), %rdi
|
|
1986
|
+
leaq TABENT(%rsp), %rsi
|
|
1987
|
+
leaq BTABENT(%rsp), %rbp
|
|
1988
|
+
callq edwards25519_scalarmuldouble_pepadd
|
|
1989
|
+
|
|
1990
|
+
// Double to acc' = 8 * acc
|
|
1991
|
+
|
|
1992
|
+
leaq ACC(%rsp), %rdi
|
|
1993
|
+
leaq ACC(%rsp), %rsi
|
|
1994
|
+
callq edwards25519_scalarmuldouble_pdouble
|
|
1995
|
+
|
|
1996
|
+
// Double to acc' = 16 * acc
|
|
1997
|
+
|
|
1998
|
+
leaq ACC(%rsp), %rdi
|
|
1999
|
+
leaq ACC(%rsp), %rsi
|
|
2000
|
+
callq edwards25519_scalarmuldouble_epdouble
|
|
2001
|
+
|
|
2002
|
+
// Add table entry, acc := acc + tabent
|
|
2003
|
+
|
|
2004
|
+
leaq ACC(%rsp), %rdi
|
|
2005
|
+
leaq ACC(%rsp), %rsi
|
|
2006
|
+
leaq TABENT(%rsp), %rbp
|
|
2007
|
+
callq edwards25519_scalarmuldouble_epadd
|
|
2008
|
+
|
|
2009
|
+
// Loop down
|
|
2010
|
+
|
|
2011
|
+
movq i, %rax
|
|
2012
|
+
testq %rax, %rax
|
|
2013
|
+
jnz edwards25519_scalarmuldouble_loop
|
|
2014
|
+
|
|
2015
|
+
// Prepare to call the modular inverse function to get tab = 1/z
|
|
2016
|
+
|
|
2017
|
+
leaq TAB(%rsp), %rdi
|
|
2018
|
+
leaq ACC+64(%rsp), %rsi
|
|
2019
|
+
|
|
2020
|
+
// Inline copy of bignum_inv_p25519, identical except for stripping out
|
|
2021
|
+
// the prologue and epilogue saving and restoring registers and making
|
|
2022
|
+
// and reclaiming room on the stack. For more details and explanations see
|
|
2023
|
+
// "x86/curve25519/bignum_inv_p25519.S". Note that the stack it uses for
|
|
2024
|
+
// its own temporaries is 208 bytes, so it has no effect on variables
|
|
2025
|
+
// that are needed in the rest of our computation here: res, tab and acc.
|
|
2026
|
+
|
|
2027
|
+
movq %rdi, 0xc0(%rsp)
|
|
2028
|
+
xorl %eax, %eax
|
|
2029
|
+
leaq -0x13(%rax), %rcx
|
|
2030
|
+
notq %rax
|
|
2031
|
+
movq %rcx, (%rsp)
|
|
2032
|
+
movq %rax, 0x8(%rsp)
|
|
2033
|
+
movq %rax, 0x10(%rsp)
|
|
2034
|
+
btr $0x3f, %rax
|
|
2035
|
+
movq %rax, 0x18(%rsp)
|
|
2036
|
+
movq (%rsi), %rdx
|
|
2037
|
+
movq 0x8(%rsi), %rcx
|
|
2038
|
+
movq 0x10(%rsi), %r8
|
|
2039
|
+
movq 0x18(%rsi), %r9
|
|
2040
|
+
movl $0x1, %eax
|
|
2041
|
+
xorl %r10d, %r10d
|
|
2042
|
+
bts $0x3f, %r9
|
|
2043
|
+
adcq %r10, %rax
|
|
2044
|
+
imulq $0x13, %rax, %rax
|
|
2045
|
+
addq %rax, %rdx
|
|
2046
|
+
adcq %r10, %rcx
|
|
2047
|
+
adcq %r10, %r8
|
|
2048
|
+
adcq %r10, %r9
|
|
2049
|
+
movl $0x13, %eax
|
|
2050
|
+
cmovbq %r10, %rax
|
|
2051
|
+
subq %rax, %rdx
|
|
2052
|
+
sbbq %r10, %rcx
|
|
2053
|
+
sbbq %r10, %r8
|
|
2054
|
+
sbbq %r10, %r9
|
|
2055
|
+
btr $0x3f, %r9
|
|
2056
|
+
movq %rdx, 0x20(%rsp)
|
|
2057
|
+
movq %rcx, 0x28(%rsp)
|
|
2058
|
+
movq %r8, 0x30(%rsp)
|
|
2059
|
+
movq %r9, 0x38(%rsp)
|
|
2060
|
+
xorl %eax, %eax
|
|
2061
|
+
movq %rax, 0x40(%rsp)
|
|
2062
|
+
movq %rax, 0x48(%rsp)
|
|
2063
|
+
movq %rax, 0x50(%rsp)
|
|
2064
|
+
movq %rax, 0x58(%rsp)
|
|
2065
|
+
movabsq $0xa0f99e2375022099, %rax
|
|
2066
|
+
movq %rax, 0x60(%rsp)
|
|
2067
|
+
movabsq $0xa8c68f3f1d132595, %rax
|
|
2068
|
+
movq %rax, 0x68(%rsp)
|
|
2069
|
+
movabsq $0x6c6c893805ac5242, %rax
|
|
2070
|
+
movq %rax, 0x70(%rsp)
|
|
2071
|
+
movabsq $0x276508b241770615, %rax
|
|
2072
|
+
movq %rax, 0x78(%rsp)
|
|
2073
|
+
movq $0xa, 0x90(%rsp)
|
|
2074
|
+
movq $0x1, 0x98(%rsp)
|
|
2075
|
+
jmp edwards25519_scalarmuldouble_midloop
|
|
2076
|
+
edwards25519_scalarmuldouble_inverseloop:
|
|
2077
|
+
movq %r8, %r9
|
|
2078
|
+
sarq $0x3f, %r9
|
|
2079
|
+
xorq %r9, %r8
|
|
2080
|
+
subq %r9, %r8
|
|
2081
|
+
movq %r10, %r11
|
|
2082
|
+
sarq $0x3f, %r11
|
|
2083
|
+
xorq %r11, %r10
|
|
2084
|
+
subq %r11, %r10
|
|
2085
|
+
movq %r12, %r13
|
|
2086
|
+
sarq $0x3f, %r13
|
|
2087
|
+
xorq %r13, %r12
|
|
2088
|
+
subq %r13, %r12
|
|
2089
|
+
movq %r14, %r15
|
|
2090
|
+
sarq $0x3f, %r15
|
|
2091
|
+
xorq %r15, %r14
|
|
2092
|
+
subq %r15, %r14
|
|
2093
|
+
movq %r8, %rax
|
|
2094
|
+
andq %r9, %rax
|
|
2095
|
+
movq %r10, %rdi
|
|
2096
|
+
andq %r11, %rdi
|
|
2097
|
+
addq %rax, %rdi
|
|
2098
|
+
movq %rdi, 0x80(%rsp)
|
|
2099
|
+
movq %r12, %rax
|
|
2100
|
+
andq %r13, %rax
|
|
2101
|
+
movq %r14, %rsi
|
|
2102
|
+
andq %r15, %rsi
|
|
2103
|
+
addq %rax, %rsi
|
|
2104
|
+
movq %rsi, 0x88(%rsp)
|
|
2105
|
+
xorl %ebx, %ebx
|
|
2106
|
+
movq (%rsp), %rax
|
|
2107
|
+
xorq %r9, %rax
|
|
2108
|
+
mulq %r8
|
|
2109
|
+
addq %rax, %rdi
|
|
2110
|
+
adcq %rdx, %rbx
|
|
2111
|
+
movq 0x20(%rsp), %rax
|
|
2112
|
+
xorq %r11, %rax
|
|
2113
|
+
mulq %r10
|
|
2114
|
+
addq %rax, %rdi
|
|
2115
|
+
adcq %rdx, %rbx
|
|
2116
|
+
xorl %ebp, %ebp
|
|
2117
|
+
movq (%rsp), %rax
|
|
2118
|
+
xorq %r13, %rax
|
|
2119
|
+
mulq %r12
|
|
2120
|
+
addq %rax, %rsi
|
|
2121
|
+
adcq %rdx, %rbp
|
|
2122
|
+
movq 0x20(%rsp), %rax
|
|
2123
|
+
xorq %r15, %rax
|
|
2124
|
+
mulq %r14
|
|
2125
|
+
addq %rax, %rsi
|
|
2126
|
+
adcq %rdx, %rbp
|
|
2127
|
+
xorl %ecx, %ecx
|
|
2128
|
+
movq 0x8(%rsp), %rax
|
|
2129
|
+
xorq %r9, %rax
|
|
2130
|
+
mulq %r8
|
|
2131
|
+
addq %rax, %rbx
|
|
2132
|
+
adcq %rdx, %rcx
|
|
2133
|
+
movq 0x28(%rsp), %rax
|
|
2134
|
+
xorq %r11, %rax
|
|
2135
|
+
mulq %r10
|
|
2136
|
+
addq %rax, %rbx
|
|
2137
|
+
adcq %rdx, %rcx
|
|
2138
|
+
shrdq $0x3b, %rbx, %rdi
|
|
2139
|
+
movq %rdi, (%rsp)
|
|
2140
|
+
xorl %edi, %edi
|
|
2141
|
+
movq 0x8(%rsp), %rax
|
|
2142
|
+
xorq %r13, %rax
|
|
2143
|
+
mulq %r12
|
|
2144
|
+
addq %rax, %rbp
|
|
2145
|
+
adcq %rdx, %rdi
|
|
2146
|
+
movq 0x28(%rsp), %rax
|
|
2147
|
+
xorq %r15, %rax
|
|
2148
|
+
mulq %r14
|
|
2149
|
+
addq %rax, %rbp
|
|
2150
|
+
adcq %rdx, %rdi
|
|
2151
|
+
shrdq $0x3b, %rbp, %rsi
|
|
2152
|
+
movq %rsi, 0x20(%rsp)
|
|
2153
|
+
xorl %esi, %esi
|
|
2154
|
+
movq 0x10(%rsp), %rax
|
|
2155
|
+
xorq %r9, %rax
|
|
2156
|
+
mulq %r8
|
|
2157
|
+
addq %rax, %rcx
|
|
2158
|
+
adcq %rdx, %rsi
|
|
2159
|
+
movq 0x30(%rsp), %rax
|
|
2160
|
+
xorq %r11, %rax
|
|
2161
|
+
mulq %r10
|
|
2162
|
+
addq %rax, %rcx
|
|
2163
|
+
adcq %rdx, %rsi
|
|
2164
|
+
shrdq $0x3b, %rcx, %rbx
|
|
2165
|
+
movq %rbx, 0x8(%rsp)
|
|
2166
|
+
xorl %ebx, %ebx
|
|
2167
|
+
movq 0x10(%rsp), %rax
|
|
2168
|
+
xorq %r13, %rax
|
|
2169
|
+
mulq %r12
|
|
2170
|
+
addq %rax, %rdi
|
|
2171
|
+
adcq %rdx, %rbx
|
|
2172
|
+
movq 0x30(%rsp), %rax
|
|
2173
|
+
xorq %r15, %rax
|
|
2174
|
+
mulq %r14
|
|
2175
|
+
addq %rax, %rdi
|
|
2176
|
+
adcq %rdx, %rbx
|
|
2177
|
+
shrdq $0x3b, %rdi, %rbp
|
|
2178
|
+
movq %rbp, 0x28(%rsp)
|
|
2179
|
+
movq 0x18(%rsp), %rax
|
|
2180
|
+
xorq %r9, %rax
|
|
2181
|
+
movq %rax, %rbp
|
|
2182
|
+
sarq $0x3f, %rbp
|
|
2183
|
+
andq %r8, %rbp
|
|
2184
|
+
negq %rbp
|
|
2185
|
+
mulq %r8
|
|
2186
|
+
addq %rax, %rsi
|
|
2187
|
+
adcq %rdx, %rbp
|
|
2188
|
+
movq 0x38(%rsp), %rax
|
|
2189
|
+
xorq %r11, %rax
|
|
2190
|
+
movq %rax, %rdx
|
|
2191
|
+
sarq $0x3f, %rdx
|
|
2192
|
+
andq %r10, %rdx
|
|
2193
|
+
subq %rdx, %rbp
|
|
2194
|
+
mulq %r10
|
|
2195
|
+
addq %rax, %rsi
|
|
2196
|
+
adcq %rdx, %rbp
|
|
2197
|
+
shrdq $0x3b, %rsi, %rcx
|
|
2198
|
+
movq %rcx, 0x10(%rsp)
|
|
2199
|
+
shrdq $0x3b, %rbp, %rsi
|
|
2200
|
+
movq 0x18(%rsp), %rax
|
|
2201
|
+
movq %rsi, 0x18(%rsp)
|
|
2202
|
+
xorq %r13, %rax
|
|
2203
|
+
movq %rax, %rsi
|
|
2204
|
+
sarq $0x3f, %rsi
|
|
2205
|
+
andq %r12, %rsi
|
|
2206
|
+
negq %rsi
|
|
2207
|
+
mulq %r12
|
|
2208
|
+
addq %rax, %rbx
|
|
2209
|
+
adcq %rdx, %rsi
|
|
2210
|
+
movq 0x38(%rsp), %rax
|
|
2211
|
+
xorq %r15, %rax
|
|
2212
|
+
movq %rax, %rdx
|
|
2213
|
+
sarq $0x3f, %rdx
|
|
2214
|
+
andq %r14, %rdx
|
|
2215
|
+
subq %rdx, %rsi
|
|
2216
|
+
mulq %r14
|
|
2217
|
+
addq %rax, %rbx
|
|
2218
|
+
adcq %rdx, %rsi
|
|
2219
|
+
shrdq $0x3b, %rbx, %rdi
|
|
2220
|
+
movq %rdi, 0x30(%rsp)
|
|
2221
|
+
shrdq $0x3b, %rsi, %rbx
|
|
2222
|
+
movq %rbx, 0x38(%rsp)
|
|
2223
|
+
movq 0x80(%rsp), %rbx
|
|
2224
|
+
movq 0x88(%rsp), %rbp
|
|
2225
|
+
xorl %ecx, %ecx
|
|
2226
|
+
movq 0x40(%rsp), %rax
|
|
2227
|
+
xorq %r9, %rax
|
|
2228
|
+
mulq %r8
|
|
2229
|
+
addq %rax, %rbx
|
|
2230
|
+
adcq %rdx, %rcx
|
|
2231
|
+
movq 0x60(%rsp), %rax
|
|
2232
|
+
xorq %r11, %rax
|
|
2233
|
+
mulq %r10
|
|
2234
|
+
addq %rax, %rbx
|
|
2235
|
+
adcq %rdx, %rcx
|
|
2236
|
+
xorl %esi, %esi
|
|
2237
|
+
movq 0x40(%rsp), %rax
|
|
2238
|
+
xorq %r13, %rax
|
|
2239
|
+
mulq %r12
|
|
2240
|
+
movq %rbx, 0x40(%rsp)
|
|
2241
|
+
addq %rax, %rbp
|
|
2242
|
+
adcq %rdx, %rsi
|
|
2243
|
+
movq 0x60(%rsp), %rax
|
|
2244
|
+
xorq %r15, %rax
|
|
2245
|
+
mulq %r14
|
|
2246
|
+
addq %rax, %rbp
|
|
2247
|
+
adcq %rdx, %rsi
|
|
2248
|
+
movq %rbp, 0x60(%rsp)
|
|
2249
|
+
xorl %ebx, %ebx
|
|
2250
|
+
movq 0x48(%rsp), %rax
|
|
2251
|
+
xorq %r9, %rax
|
|
2252
|
+
mulq %r8
|
|
2253
|
+
addq %rax, %rcx
|
|
2254
|
+
adcq %rdx, %rbx
|
|
2255
|
+
movq 0x68(%rsp), %rax
|
|
2256
|
+
xorq %r11, %rax
|
|
2257
|
+
mulq %r10
|
|
2258
|
+
addq %rax, %rcx
|
|
2259
|
+
adcq %rdx, %rbx
|
|
2260
|
+
xorl %ebp, %ebp
|
|
2261
|
+
movq 0x48(%rsp), %rax
|
|
2262
|
+
xorq %r13, %rax
|
|
2263
|
+
mulq %r12
|
|
2264
|
+
movq %rcx, 0x48(%rsp)
|
|
2265
|
+
addq %rax, %rsi
|
|
2266
|
+
adcq %rdx, %rbp
|
|
2267
|
+
movq 0x68(%rsp), %rax
|
|
2268
|
+
xorq %r15, %rax
|
|
2269
|
+
mulq %r14
|
|
2270
|
+
addq %rax, %rsi
|
|
2271
|
+
adcq %rdx, %rbp
|
|
2272
|
+
movq %rsi, 0x68(%rsp)
|
|
2273
|
+
xorl %ecx, %ecx
|
|
2274
|
+
movq 0x50(%rsp), %rax
|
|
2275
|
+
xorq %r9, %rax
|
|
2276
|
+
mulq %r8
|
|
2277
|
+
addq %rax, %rbx
|
|
2278
|
+
adcq %rdx, %rcx
|
|
2279
|
+
movq 0x70(%rsp), %rax
|
|
2280
|
+
xorq %r11, %rax
|
|
2281
|
+
mulq %r10
|
|
2282
|
+
addq %rax, %rbx
|
|
2283
|
+
adcq %rdx, %rcx
|
|
2284
|
+
xorl %esi, %esi
|
|
2285
|
+
movq 0x50(%rsp), %rax
|
|
2286
|
+
xorq %r13, %rax
|
|
2287
|
+
mulq %r12
|
|
2288
|
+
movq %rbx, 0x50(%rsp)
|
|
2289
|
+
addq %rax, %rbp
|
|
2290
|
+
adcq %rdx, %rsi
|
|
2291
|
+
movq 0x70(%rsp), %rax
|
|
2292
|
+
xorq %r15, %rax
|
|
2293
|
+
mulq %r14
|
|
2294
|
+
addq %rax, %rbp
|
|
2295
|
+
adcq %rdx, %rsi
|
|
2296
|
+
movq %rbp, 0x70(%rsp)
|
|
2297
|
+
movq 0x58(%rsp), %rax
|
|
2298
|
+
xorq %r9, %rax
|
|
2299
|
+
movq %r9, %rbx
|
|
2300
|
+
andq %r8, %rbx
|
|
2301
|
+
negq %rbx
|
|
2302
|
+
mulq %r8
|
|
2303
|
+
addq %rax, %rcx
|
|
2304
|
+
adcq %rdx, %rbx
|
|
2305
|
+
movq 0x78(%rsp), %rax
|
|
2306
|
+
xorq %r11, %rax
|
|
2307
|
+
movq %r11, %rdx
|
|
2308
|
+
andq %r10, %rdx
|
|
2309
|
+
subq %rdx, %rbx
|
|
2310
|
+
mulq %r10
|
|
2311
|
+
addq %rax, %rcx
|
|
2312
|
+
adcq %rbx, %rdx
|
|
2313
|
+
movq %rdx, %rbx
|
|
2314
|
+
shldq $0x1, %rcx, %rdx
|
|
2315
|
+
sarq $0x3f, %rbx
|
|
2316
|
+
addq %rbx, %rdx
|
|
2317
|
+
movl $0x13, %eax
|
|
2318
|
+
imulq %rdx
|
|
2319
|
+
movq 0x40(%rsp), %r8
|
|
2320
|
+
addq %rax, %r8
|
|
2321
|
+
movq %r8, 0x40(%rsp)
|
|
2322
|
+
movq 0x48(%rsp), %r8
|
|
2323
|
+
adcq %rdx, %r8
|
|
2324
|
+
movq %r8, 0x48(%rsp)
|
|
2325
|
+
movq 0x50(%rsp), %r8
|
|
2326
|
+
adcq %rbx, %r8
|
|
2327
|
+
movq %r8, 0x50(%rsp)
|
|
2328
|
+
adcq %rbx, %rcx
|
|
2329
|
+
shlq $0x3f, %rax
|
|
2330
|
+
addq %rax, %rcx
|
|
2331
|
+
movq 0x58(%rsp), %rax
|
|
2332
|
+
movq %rcx, 0x58(%rsp)
|
|
2333
|
+
xorq %r13, %rax
|
|
2334
|
+
movq %r13, %rcx
|
|
2335
|
+
andq %r12, %rcx
|
|
2336
|
+
negq %rcx
|
|
2337
|
+
mulq %r12
|
|
2338
|
+
addq %rax, %rsi
|
|
2339
|
+
adcq %rdx, %rcx
|
|
2340
|
+
movq 0x78(%rsp), %rax
|
|
2341
|
+
xorq %r15, %rax
|
|
2342
|
+
movq %r15, %rdx
|
|
2343
|
+
andq %r14, %rdx
|
|
2344
|
+
subq %rdx, %rcx
|
|
2345
|
+
mulq %r14
|
|
2346
|
+
addq %rax, %rsi
|
|
2347
|
+
adcq %rcx, %rdx
|
|
2348
|
+
movq %rdx, %rcx
|
|
2349
|
+
shldq $0x1, %rsi, %rdx
|
|
2350
|
+
sarq $0x3f, %rcx
|
|
2351
|
+
movl $0x13, %eax
|
|
2352
|
+
addq %rcx, %rdx
|
|
2353
|
+
imulq %rdx
|
|
2354
|
+
movq 0x60(%rsp), %r8
|
|
2355
|
+
addq %rax, %r8
|
|
2356
|
+
movq %r8, 0x60(%rsp)
|
|
2357
|
+
movq 0x68(%rsp), %r8
|
|
2358
|
+
adcq %rdx, %r8
|
|
2359
|
+
movq %r8, 0x68(%rsp)
|
|
2360
|
+
movq 0x70(%rsp), %r8
|
|
2361
|
+
adcq %rcx, %r8
|
|
2362
|
+
movq %r8, 0x70(%rsp)
|
|
2363
|
+
adcq %rcx, %rsi
|
|
2364
|
+
shlq $0x3f, %rax
|
|
2365
|
+
addq %rax, %rsi
|
|
2366
|
+
movq %rsi, 0x78(%rsp)
|
|
2367
|
+
edwards25519_scalarmuldouble_midloop:
|
|
2368
|
+
movq 0x98(%rsp), %rsi
|
|
2369
|
+
movq (%rsp), %rdx
|
|
2370
|
+
movq 0x20(%rsp), %rcx
|
|
2371
|
+
movq %rdx, %rbx
|
|
2372
|
+
andq $0xfffff, %rbx
|
|
2373
|
+
movabsq $0xfffffe0000000000, %rax
|
|
2374
|
+
orq %rax, %rbx
|
|
2375
|
+
andq $0xfffff, %rcx
|
|
2376
|
+
movabsq $0xc000000000000000, %rax
|
|
2377
|
+
orq %rax, %rcx
|
|
2378
|
+
movq $0xfffffffffffffffe, %rax
|
|
2379
|
+
xorl %ebp, %ebp
|
|
2380
|
+
movl $0x2, %edx
|
|
2381
|
+
movq %rbx, %rdi
|
|
2382
|
+
movq %rax, %r8
|
|
2383
|
+
testq %rsi, %rsi
|
|
2384
|
+
cmovs %rbp, %r8
|
|
2385
|
+
testq $0x1, %rcx
|
|
2386
|
+
cmoveq %rbp, %r8
|
|
2387
|
+
cmoveq %rbp, %rdi
|
|
2388
|
+
xorq %r8, %rdi
|
|
2389
|
+
xorq %r8, %rsi
|
|
2390
|
+
btq $0x3f, %r8
|
|
2391
|
+
cmovbq %rcx, %rbx
|
|
2392
|
+
movq %rax, %r8
|
|
2393
|
+
subq %rax, %rsi
|
|
2394
|
+
leaq (%rcx,%rdi), %rcx
|
|
2395
|
+
cmovs %rbp, %r8
|
|
2396
|
+
movq %rbx, %rdi
|
|
2397
|
+
testq %rdx, %rcx
|
|
2398
|
+
cmoveq %rbp, %r8
|
|
2399
|
+
cmoveq %rbp, %rdi
|
|
2400
|
+
sarq $1, %rcx
|
|
2401
|
+
xorq %r8, %rdi
|
|
2402
|
+
xorq %r8, %rsi
|
|
2403
|
+
btq $0x3f, %r8
|
|
2404
|
+
cmovbq %rcx, %rbx
|
|
2405
|
+
movq %rax, %r8
|
|
2406
|
+
subq %rax, %rsi
|
|
2407
|
+
leaq (%rcx,%rdi), %rcx
|
|
2408
|
+
cmovs %rbp, %r8
|
|
2409
|
+
movq %rbx, %rdi
|
|
2410
|
+
testq %rdx, %rcx
|
|
2411
|
+
cmoveq %rbp, %r8
|
|
2412
|
+
cmoveq %rbp, %rdi
|
|
2413
|
+
sarq $1, %rcx
|
|
2414
|
+
xorq %r8, %rdi
|
|
2415
|
+
xorq %r8, %rsi
|
|
2416
|
+
btq $0x3f, %r8
|
|
2417
|
+
cmovbq %rcx, %rbx
|
|
2418
|
+
movq %rax, %r8
|
|
2419
|
+
subq %rax, %rsi
|
|
2420
|
+
leaq (%rcx,%rdi), %rcx
|
|
2421
|
+
cmovs %rbp, %r8
|
|
2422
|
+
movq %rbx, %rdi
|
|
2423
|
+
testq %rdx, %rcx
|
|
2424
|
+
cmoveq %rbp, %r8
|
|
2425
|
+
cmoveq %rbp, %rdi
|
|
2426
|
+
sarq $1, %rcx
|
|
2427
|
+
xorq %r8, %rdi
|
|
2428
|
+
xorq %r8, %rsi
|
|
2429
|
+
btq $0x3f, %r8
|
|
2430
|
+
cmovbq %rcx, %rbx
|
|
2431
|
+
movq %rax, %r8
|
|
2432
|
+
subq %rax, %rsi
|
|
2433
|
+
leaq (%rcx,%rdi), %rcx
|
|
2434
|
+
cmovs %rbp, %r8
|
|
2435
|
+
movq %rbx, %rdi
|
|
2436
|
+
testq %rdx, %rcx
|
|
2437
|
+
cmoveq %rbp, %r8
|
|
2438
|
+
cmoveq %rbp, %rdi
|
|
2439
|
+
sarq $1, %rcx
|
|
2440
|
+
xorq %r8, %rdi
|
|
2441
|
+
xorq %r8, %rsi
|
|
2442
|
+
btq $0x3f, %r8
|
|
2443
|
+
cmovbq %rcx, %rbx
|
|
2444
|
+
movq %rax, %r8
|
|
2445
|
+
subq %rax, %rsi
|
|
2446
|
+
leaq (%rcx,%rdi), %rcx
|
|
2447
|
+
cmovs %rbp, %r8
|
|
2448
|
+
movq %rbx, %rdi
|
|
2449
|
+
testq %rdx, %rcx
|
|
2450
|
+
cmoveq %rbp, %r8
|
|
2451
|
+
cmoveq %rbp, %rdi
|
|
2452
|
+
sarq $1, %rcx
|
|
2453
|
+
xorq %r8, %rdi
|
|
2454
|
+
xorq %r8, %rsi
|
|
2455
|
+
btq $0x3f, %r8
|
|
2456
|
+
cmovbq %rcx, %rbx
|
|
2457
|
+
movq %rax, %r8
|
|
2458
|
+
subq %rax, %rsi
|
|
2459
|
+
leaq (%rcx,%rdi), %rcx
|
|
2460
|
+
cmovs %rbp, %r8
|
|
2461
|
+
movq %rbx, %rdi
|
|
2462
|
+
testq %rdx, %rcx
|
|
2463
|
+
cmoveq %rbp, %r8
|
|
2464
|
+
cmoveq %rbp, %rdi
|
|
2465
|
+
sarq $1, %rcx
|
|
2466
|
+
xorq %r8, %rdi
|
|
2467
|
+
xorq %r8, %rsi
|
|
2468
|
+
btq $0x3f, %r8
|
|
2469
|
+
cmovbq %rcx, %rbx
|
|
2470
|
+
movq %rax, %r8
|
|
2471
|
+
subq %rax, %rsi
|
|
2472
|
+
leaq (%rcx,%rdi), %rcx
|
|
2473
|
+
cmovs %rbp, %r8
|
|
2474
|
+
movq %rbx, %rdi
|
|
2475
|
+
testq %rdx, %rcx
|
|
2476
|
+
cmoveq %rbp, %r8
|
|
2477
|
+
cmoveq %rbp, %rdi
|
|
2478
|
+
sarq $1, %rcx
|
|
2479
|
+
xorq %r8, %rdi
|
|
2480
|
+
xorq %r8, %rsi
|
|
2481
|
+
btq $0x3f, %r8
|
|
2482
|
+
cmovbq %rcx, %rbx
|
|
2483
|
+
movq %rax, %r8
|
|
2484
|
+
subq %rax, %rsi
|
|
2485
|
+
leaq (%rcx,%rdi), %rcx
|
|
2486
|
+
cmovs %rbp, %r8
|
|
2487
|
+
movq %rbx, %rdi
|
|
2488
|
+
testq %rdx, %rcx
|
|
2489
|
+
cmoveq %rbp, %r8
|
|
2490
|
+
cmoveq %rbp, %rdi
|
|
2491
|
+
sarq $1, %rcx
|
|
2492
|
+
xorq %r8, %rdi
|
|
2493
|
+
xorq %r8, %rsi
|
|
2494
|
+
btq $0x3f, %r8
|
|
2495
|
+
cmovbq %rcx, %rbx
|
|
2496
|
+
movq %rax, %r8
|
|
2497
|
+
subq %rax, %rsi
|
|
2498
|
+
leaq (%rcx,%rdi), %rcx
|
|
2499
|
+
cmovs %rbp, %r8
|
|
2500
|
+
movq %rbx, %rdi
|
|
2501
|
+
testq %rdx, %rcx
|
|
2502
|
+
cmoveq %rbp, %r8
|
|
2503
|
+
cmoveq %rbp, %rdi
|
|
2504
|
+
sarq $1, %rcx
|
|
2505
|
+
xorq %r8, %rdi
|
|
2506
|
+
xorq %r8, %rsi
|
|
2507
|
+
btq $0x3f, %r8
|
|
2508
|
+
cmovbq %rcx, %rbx
|
|
2509
|
+
movq %rax, %r8
|
|
2510
|
+
subq %rax, %rsi
|
|
2511
|
+
leaq (%rcx,%rdi), %rcx
|
|
2512
|
+
cmovs %rbp, %r8
|
|
2513
|
+
movq %rbx, %rdi
|
|
2514
|
+
testq %rdx, %rcx
|
|
2515
|
+
cmoveq %rbp, %r8
|
|
2516
|
+
cmoveq %rbp, %rdi
|
|
2517
|
+
sarq $1, %rcx
|
|
2518
|
+
xorq %r8, %rdi
|
|
2519
|
+
xorq %r8, %rsi
|
|
2520
|
+
btq $0x3f, %r8
|
|
2521
|
+
cmovbq %rcx, %rbx
|
|
2522
|
+
movq %rax, %r8
|
|
2523
|
+
subq %rax, %rsi
|
|
2524
|
+
leaq (%rcx,%rdi), %rcx
|
|
2525
|
+
cmovs %rbp, %r8
|
|
2526
|
+
movq %rbx, %rdi
|
|
2527
|
+
testq %rdx, %rcx
|
|
2528
|
+
cmoveq %rbp, %r8
|
|
2529
|
+
cmoveq %rbp, %rdi
|
|
2530
|
+
sarq $1, %rcx
|
|
2531
|
+
xorq %r8, %rdi
|
|
2532
|
+
xorq %r8, %rsi
|
|
2533
|
+
btq $0x3f, %r8
|
|
2534
|
+
cmovbq %rcx, %rbx
|
|
2535
|
+
movq %rax, %r8
|
|
2536
|
+
subq %rax, %rsi
|
|
2537
|
+
leaq (%rcx,%rdi), %rcx
|
|
2538
|
+
cmovs %rbp, %r8
|
|
2539
|
+
movq %rbx, %rdi
|
|
2540
|
+
testq %rdx, %rcx
|
|
2541
|
+
cmoveq %rbp, %r8
|
|
2542
|
+
cmoveq %rbp, %rdi
|
|
2543
|
+
sarq $1, %rcx
|
|
2544
|
+
xorq %r8, %rdi
|
|
2545
|
+
xorq %r8, %rsi
|
|
2546
|
+
btq $0x3f, %r8
|
|
2547
|
+
cmovbq %rcx, %rbx
|
|
2548
|
+
movq %rax, %r8
|
|
2549
|
+
subq %rax, %rsi
|
|
2550
|
+
leaq (%rcx,%rdi), %rcx
|
|
2551
|
+
cmovs %rbp, %r8
|
|
2552
|
+
movq %rbx, %rdi
|
|
2553
|
+
testq %rdx, %rcx
|
|
2554
|
+
cmoveq %rbp, %r8
|
|
2555
|
+
cmoveq %rbp, %rdi
|
|
2556
|
+
sarq $1, %rcx
|
|
2557
|
+
xorq %r8, %rdi
|
|
2558
|
+
xorq %r8, %rsi
|
|
2559
|
+
btq $0x3f, %r8
|
|
2560
|
+
cmovbq %rcx, %rbx
|
|
2561
|
+
movq %rax, %r8
|
|
2562
|
+
subq %rax, %rsi
|
|
2563
|
+
leaq (%rcx,%rdi), %rcx
|
|
2564
|
+
cmovs %rbp, %r8
|
|
2565
|
+
movq %rbx, %rdi
|
|
2566
|
+
testq %rdx, %rcx
|
|
2567
|
+
cmoveq %rbp, %r8
|
|
2568
|
+
cmoveq %rbp, %rdi
|
|
2569
|
+
sarq $1, %rcx
|
|
2570
|
+
xorq %r8, %rdi
|
|
2571
|
+
xorq %r8, %rsi
|
|
2572
|
+
btq $0x3f, %r8
|
|
2573
|
+
cmovbq %rcx, %rbx
|
|
2574
|
+
movq %rax, %r8
|
|
2575
|
+
subq %rax, %rsi
|
|
2576
|
+
leaq (%rcx,%rdi), %rcx
|
|
2577
|
+
cmovs %rbp, %r8
|
|
2578
|
+
movq %rbx, %rdi
|
|
2579
|
+
testq %rdx, %rcx
|
|
2580
|
+
cmoveq %rbp, %r8
|
|
2581
|
+
cmoveq %rbp, %rdi
|
|
2582
|
+
sarq $1, %rcx
|
|
2583
|
+
xorq %r8, %rdi
|
|
2584
|
+
xorq %r8, %rsi
|
|
2585
|
+
btq $0x3f, %r8
|
|
2586
|
+
cmovbq %rcx, %rbx
|
|
2587
|
+
movq %rax, %r8
|
|
2588
|
+
subq %rax, %rsi
|
|
2589
|
+
leaq (%rcx,%rdi), %rcx
|
|
2590
|
+
cmovs %rbp, %r8
|
|
2591
|
+
movq %rbx, %rdi
|
|
2592
|
+
testq %rdx, %rcx
|
|
2593
|
+
cmoveq %rbp, %r8
|
|
2594
|
+
cmoveq %rbp, %rdi
|
|
2595
|
+
sarq $1, %rcx
|
|
2596
|
+
xorq %r8, %rdi
|
|
2597
|
+
xorq %r8, %rsi
|
|
2598
|
+
btq $0x3f, %r8
|
|
2599
|
+
cmovbq %rcx, %rbx
|
|
2600
|
+
movq %rax, %r8
|
|
2601
|
+
subq %rax, %rsi
|
|
2602
|
+
leaq (%rcx,%rdi), %rcx
|
|
2603
|
+
cmovs %rbp, %r8
|
|
2604
|
+
movq %rbx, %rdi
|
|
2605
|
+
testq %rdx, %rcx
|
|
2606
|
+
cmoveq %rbp, %r8
|
|
2607
|
+
cmoveq %rbp, %rdi
|
|
2608
|
+
sarq $1, %rcx
|
|
2609
|
+
xorq %r8, %rdi
|
|
2610
|
+
xorq %r8, %rsi
|
|
2611
|
+
btq $0x3f, %r8
|
|
2612
|
+
cmovbq %rcx, %rbx
|
|
2613
|
+
movq %rax, %r8
|
|
2614
|
+
subq %rax, %rsi
|
|
2615
|
+
leaq (%rcx,%rdi), %rcx
|
|
2616
|
+
cmovs %rbp, %r8
|
|
2617
|
+
movq %rbx, %rdi
|
|
2618
|
+
testq %rdx, %rcx
|
|
2619
|
+
cmoveq %rbp, %r8
|
|
2620
|
+
cmoveq %rbp, %rdi
|
|
2621
|
+
sarq $1, %rcx
|
|
2622
|
+
xorq %r8, %rdi
|
|
2623
|
+
xorq %r8, %rsi
|
|
2624
|
+
btq $0x3f, %r8
|
|
2625
|
+
cmovbq %rcx, %rbx
|
|
2626
|
+
movq %rax, %r8
|
|
2627
|
+
subq %rax, %rsi
|
|
2628
|
+
leaq (%rcx,%rdi), %rcx
|
|
2629
|
+
cmovs %rbp, %r8
|
|
2630
|
+
movq %rbx, %rdi
|
|
2631
|
+
testq %rdx, %rcx
|
|
2632
|
+
cmoveq %rbp, %r8
|
|
2633
|
+
cmoveq %rbp, %rdi
|
|
2634
|
+
sarq $1, %rcx
|
|
2635
|
+
xorq %r8, %rdi
|
|
2636
|
+
xorq %r8, %rsi
|
|
2637
|
+
btq $0x3f, %r8
|
|
2638
|
+
cmovbq %rcx, %rbx
|
|
2639
|
+
movq %rax, %r8
|
|
2640
|
+
subq %rax, %rsi
|
|
2641
|
+
leaq (%rcx,%rdi), %rcx
|
|
2642
|
+
sarq $1, %rcx
|
|
2643
|
+
movl $0x100000, %eax
|
|
2644
|
+
leaq (%rbx,%rax), %rdx
|
|
2645
|
+
leaq (%rcx,%rax), %rdi
|
|
2646
|
+
shlq $0x16, %rdx
|
|
2647
|
+
shlq $0x16, %rdi
|
|
2648
|
+
sarq $0x2b, %rdx
|
|
2649
|
+
sarq $0x2b, %rdi
|
|
2650
|
+
movabsq $0x20000100000, %rax
|
|
2651
|
+
leaq (%rbx,%rax), %rbx
|
|
2652
|
+
leaq (%rcx,%rax), %rcx
|
|
2653
|
+
sarq $0x2a, %rbx
|
|
2654
|
+
sarq $0x2a, %rcx
|
|
2655
|
+
movq %rdx, 0xa0(%rsp)
|
|
2656
|
+
movq %rbx, 0xa8(%rsp)
|
|
2657
|
+
movq %rdi, 0xb0(%rsp)
|
|
2658
|
+
movq %rcx, 0xb8(%rsp)
|
|
2659
|
+
movq (%rsp), %r12
|
|
2660
|
+
imulq %r12, %rdi
|
|
2661
|
+
imulq %rdx, %r12
|
|
2662
|
+
movq 0x20(%rsp), %r13
|
|
2663
|
+
imulq %r13, %rbx
|
|
2664
|
+
imulq %rcx, %r13
|
|
2665
|
+
addq %rbx, %r12
|
|
2666
|
+
addq %rdi, %r13
|
|
2667
|
+
sarq $0x14, %r12
|
|
2668
|
+
sarq $0x14, %r13
|
|
2669
|
+
movq %r12, %rbx
|
|
2670
|
+
andq $0xfffff, %rbx
|
|
2671
|
+
movabsq $0xfffffe0000000000, %rax
|
|
2672
|
+
orq %rax, %rbx
|
|
2673
|
+
movq %r13, %rcx
|
|
2674
|
+
andq $0xfffff, %rcx
|
|
2675
|
+
movabsq $0xc000000000000000, %rax
|
|
2676
|
+
orq %rax, %rcx
|
|
2677
|
+
movq $0xfffffffffffffffe, %rax
|
|
2678
|
+
movl $0x2, %edx
|
|
2679
|
+
movq %rbx, %rdi
|
|
2680
|
+
movq %rax, %r8
|
|
2681
|
+
testq %rsi, %rsi
|
|
2682
|
+
cmovs %rbp, %r8
|
|
2683
|
+
testq $0x1, %rcx
|
|
2684
|
+
cmoveq %rbp, %r8
|
|
2685
|
+
cmoveq %rbp, %rdi
|
|
2686
|
+
xorq %r8, %rdi
|
|
2687
|
+
xorq %r8, %rsi
|
|
2688
|
+
btq $0x3f, %r8
|
|
2689
|
+
cmovbq %rcx, %rbx
|
|
2690
|
+
movq %rax, %r8
|
|
2691
|
+
subq %rax, %rsi
|
|
2692
|
+
leaq (%rcx,%rdi), %rcx
|
|
2693
|
+
cmovs %rbp, %r8
|
|
2694
|
+
movq %rbx, %rdi
|
|
2695
|
+
testq %rdx, %rcx
|
|
2696
|
+
cmoveq %rbp, %r8
|
|
2697
|
+
cmoveq %rbp, %rdi
|
|
2698
|
+
sarq $1, %rcx
|
|
2699
|
+
xorq %r8, %rdi
|
|
2700
|
+
xorq %r8, %rsi
|
|
2701
|
+
btq $0x3f, %r8
|
|
2702
|
+
cmovbq %rcx, %rbx
|
|
2703
|
+
movq %rax, %r8
|
|
2704
|
+
subq %rax, %rsi
|
|
2705
|
+
leaq (%rcx,%rdi), %rcx
|
|
2706
|
+
cmovs %rbp, %r8
|
|
2707
|
+
movq %rbx, %rdi
|
|
2708
|
+
testq %rdx, %rcx
|
|
2709
|
+
cmoveq %rbp, %r8
|
|
2710
|
+
cmoveq %rbp, %rdi
|
|
2711
|
+
sarq $1, %rcx
|
|
2712
|
+
xorq %r8, %rdi
|
|
2713
|
+
xorq %r8, %rsi
|
|
2714
|
+
btq $0x3f, %r8
|
|
2715
|
+
cmovbq %rcx, %rbx
|
|
2716
|
+
movq %rax, %r8
|
|
2717
|
+
subq %rax, %rsi
|
|
2718
|
+
leaq (%rcx,%rdi), %rcx
|
|
2719
|
+
cmovs %rbp, %r8
|
|
2720
|
+
movq %rbx, %rdi
|
|
2721
|
+
testq %rdx, %rcx
|
|
2722
|
+
cmoveq %rbp, %r8
|
|
2723
|
+
cmoveq %rbp, %rdi
|
|
2724
|
+
sarq $1, %rcx
|
|
2725
|
+
xorq %r8, %rdi
|
|
2726
|
+
xorq %r8, %rsi
|
|
2727
|
+
btq $0x3f, %r8
|
|
2728
|
+
cmovbq %rcx, %rbx
|
|
2729
|
+
movq %rax, %r8
|
|
2730
|
+
subq %rax, %rsi
|
|
2731
|
+
leaq (%rcx,%rdi), %rcx
|
|
2732
|
+
cmovs %rbp, %r8
|
|
2733
|
+
movq %rbx, %rdi
|
|
2734
|
+
testq %rdx, %rcx
|
|
2735
|
+
cmoveq %rbp, %r8
|
|
2736
|
+
cmoveq %rbp, %rdi
|
|
2737
|
+
sarq $1, %rcx
|
|
2738
|
+
xorq %r8, %rdi
|
|
2739
|
+
xorq %r8, %rsi
|
|
2740
|
+
btq $0x3f, %r8
|
|
2741
|
+
cmovbq %rcx, %rbx
|
|
2742
|
+
movq %rax, %r8
|
|
2743
|
+
subq %rax, %rsi
|
|
2744
|
+
leaq (%rcx,%rdi), %rcx
|
|
2745
|
+
cmovs %rbp, %r8
|
|
2746
|
+
movq %rbx, %rdi
|
|
2747
|
+
testq %rdx, %rcx
|
|
2748
|
+
cmoveq %rbp, %r8
|
|
2749
|
+
cmoveq %rbp, %rdi
|
|
2750
|
+
sarq $1, %rcx
|
|
2751
|
+
xorq %r8, %rdi
|
|
2752
|
+
xorq %r8, %rsi
|
|
2753
|
+
btq $0x3f, %r8
|
|
2754
|
+
cmovbq %rcx, %rbx
|
|
2755
|
+
movq %rax, %r8
|
|
2756
|
+
subq %rax, %rsi
|
|
2757
|
+
leaq (%rcx,%rdi), %rcx
|
|
2758
|
+
cmovs %rbp, %r8
|
|
2759
|
+
movq %rbx, %rdi
|
|
2760
|
+
testq %rdx, %rcx
|
|
2761
|
+
cmoveq %rbp, %r8
|
|
2762
|
+
cmoveq %rbp, %rdi
|
|
2763
|
+
sarq $1, %rcx
|
|
2764
|
+
xorq %r8, %rdi
|
|
2765
|
+
xorq %r8, %rsi
|
|
2766
|
+
btq $0x3f, %r8
|
|
2767
|
+
cmovbq %rcx, %rbx
|
|
2768
|
+
movq %rax, %r8
|
|
2769
|
+
subq %rax, %rsi
|
|
2770
|
+
leaq (%rcx,%rdi), %rcx
|
|
2771
|
+
cmovs %rbp, %r8
|
|
2772
|
+
movq %rbx, %rdi
|
|
2773
|
+
testq %rdx, %rcx
|
|
2774
|
+
cmoveq %rbp, %r8
|
|
2775
|
+
cmoveq %rbp, %rdi
|
|
2776
|
+
sarq $1, %rcx
|
|
2777
|
+
xorq %r8, %rdi
|
|
2778
|
+
xorq %r8, %rsi
|
|
2779
|
+
btq $0x3f, %r8
|
|
2780
|
+
cmovbq %rcx, %rbx
|
|
2781
|
+
movq %rax, %r8
|
|
2782
|
+
subq %rax, %rsi
|
|
2783
|
+
leaq (%rcx,%rdi), %rcx
|
|
2784
|
+
cmovs %rbp, %r8
|
|
2785
|
+
movq %rbx, %rdi
|
|
2786
|
+
testq %rdx, %rcx
|
|
2787
|
+
cmoveq %rbp, %r8
|
|
2788
|
+
cmoveq %rbp, %rdi
|
|
2789
|
+
sarq $1, %rcx
|
|
2790
|
+
xorq %r8, %rdi
|
|
2791
|
+
xorq %r8, %rsi
|
|
2792
|
+
btq $0x3f, %r8
|
|
2793
|
+
cmovbq %rcx, %rbx
|
|
2794
|
+
movq %rax, %r8
|
|
2795
|
+
subq %rax, %rsi
|
|
2796
|
+
leaq (%rcx,%rdi), %rcx
|
|
2797
|
+
cmovs %rbp, %r8
|
|
2798
|
+
movq %rbx, %rdi
|
|
2799
|
+
testq %rdx, %rcx
|
|
2800
|
+
cmoveq %rbp, %r8
|
|
2801
|
+
cmoveq %rbp, %rdi
|
|
2802
|
+
sarq $1, %rcx
|
|
2803
|
+
xorq %r8, %rdi
|
|
2804
|
+
xorq %r8, %rsi
|
|
2805
|
+
btq $0x3f, %r8
|
|
2806
|
+
cmovbq %rcx, %rbx
|
|
2807
|
+
movq %rax, %r8
|
|
2808
|
+
subq %rax, %rsi
|
|
2809
|
+
leaq (%rcx,%rdi), %rcx
|
|
2810
|
+
cmovs %rbp, %r8
|
|
2811
|
+
movq %rbx, %rdi
|
|
2812
|
+
testq %rdx, %rcx
|
|
2813
|
+
cmoveq %rbp, %r8
|
|
2814
|
+
cmoveq %rbp, %rdi
|
|
2815
|
+
sarq $1, %rcx
|
|
2816
|
+
xorq %r8, %rdi
|
|
2817
|
+
xorq %r8, %rsi
|
|
2818
|
+
btq $0x3f, %r8
|
|
2819
|
+
cmovbq %rcx, %rbx
|
|
2820
|
+
movq %rax, %r8
|
|
2821
|
+
subq %rax, %rsi
|
|
2822
|
+
leaq (%rcx,%rdi), %rcx
|
|
2823
|
+
cmovs %rbp, %r8
|
|
2824
|
+
movq %rbx, %rdi
|
|
2825
|
+
testq %rdx, %rcx
|
|
2826
|
+
cmoveq %rbp, %r8
|
|
2827
|
+
cmoveq %rbp, %rdi
|
|
2828
|
+
sarq $1, %rcx
|
|
2829
|
+
xorq %r8, %rdi
|
|
2830
|
+
xorq %r8, %rsi
|
|
2831
|
+
btq $0x3f, %r8
|
|
2832
|
+
cmovbq %rcx, %rbx
|
|
2833
|
+
movq %rax, %r8
|
|
2834
|
+
subq %rax, %rsi
|
|
2835
|
+
leaq (%rcx,%rdi), %rcx
|
|
2836
|
+
cmovs %rbp, %r8
|
|
2837
|
+
movq %rbx, %rdi
|
|
2838
|
+
testq %rdx, %rcx
|
|
2839
|
+
cmoveq %rbp, %r8
|
|
2840
|
+
cmoveq %rbp, %rdi
|
|
2841
|
+
sarq $1, %rcx
|
|
2842
|
+
xorq %r8, %rdi
|
|
2843
|
+
xorq %r8, %rsi
|
|
2844
|
+
btq $0x3f, %r8
|
|
2845
|
+
cmovbq %rcx, %rbx
|
|
2846
|
+
movq %rax, %r8
|
|
2847
|
+
subq %rax, %rsi
|
|
2848
|
+
leaq (%rcx,%rdi), %rcx
|
|
2849
|
+
cmovs %rbp, %r8
|
|
2850
|
+
movq %rbx, %rdi
|
|
2851
|
+
testq %rdx, %rcx
|
|
2852
|
+
cmoveq %rbp, %r8
|
|
2853
|
+
cmoveq %rbp, %rdi
|
|
2854
|
+
sarq $1, %rcx
|
|
2855
|
+
xorq %r8, %rdi
|
|
2856
|
+
xorq %r8, %rsi
|
|
2857
|
+
btq $0x3f, %r8
|
|
2858
|
+
cmovbq %rcx, %rbx
|
|
2859
|
+
movq %rax, %r8
|
|
2860
|
+
subq %rax, %rsi
|
|
2861
|
+
leaq (%rcx,%rdi), %rcx
|
|
2862
|
+
cmovs %rbp, %r8
|
|
2863
|
+
movq %rbx, %rdi
|
|
2864
|
+
testq %rdx, %rcx
|
|
2865
|
+
cmoveq %rbp, %r8
|
|
2866
|
+
cmoveq %rbp, %rdi
|
|
2867
|
+
sarq $1, %rcx
|
|
2868
|
+
xorq %r8, %rdi
|
|
2869
|
+
xorq %r8, %rsi
|
|
2870
|
+
btq $0x3f, %r8
|
|
2871
|
+
cmovbq %rcx, %rbx
|
|
2872
|
+
movq %rax, %r8
|
|
2873
|
+
subq %rax, %rsi
|
|
2874
|
+
leaq (%rcx,%rdi), %rcx
|
|
2875
|
+
cmovs %rbp, %r8
|
|
2876
|
+
movq %rbx, %rdi
|
|
2877
|
+
testq %rdx, %rcx
|
|
2878
|
+
cmoveq %rbp, %r8
|
|
2879
|
+
cmoveq %rbp, %rdi
|
|
2880
|
+
sarq $1, %rcx
|
|
2881
|
+
xorq %r8, %rdi
|
|
2882
|
+
xorq %r8, %rsi
|
|
2883
|
+
btq $0x3f, %r8
|
|
2884
|
+
cmovbq %rcx, %rbx
|
|
2885
|
+
movq %rax, %r8
|
|
2886
|
+
subq %rax, %rsi
|
|
2887
|
+
leaq (%rcx,%rdi), %rcx
|
|
2888
|
+
cmovs %rbp, %r8
|
|
2889
|
+
movq %rbx, %rdi
|
|
2890
|
+
testq %rdx, %rcx
|
|
2891
|
+
cmoveq %rbp, %r8
|
|
2892
|
+
cmoveq %rbp, %rdi
|
|
2893
|
+
sarq $1, %rcx
|
|
2894
|
+
xorq %r8, %rdi
|
|
2895
|
+
xorq %r8, %rsi
|
|
2896
|
+
btq $0x3f, %r8
|
|
2897
|
+
cmovbq %rcx, %rbx
|
|
2898
|
+
movq %rax, %r8
|
|
2899
|
+
subq %rax, %rsi
|
|
2900
|
+
leaq (%rcx,%rdi), %rcx
|
|
2901
|
+
cmovs %rbp, %r8
|
|
2902
|
+
movq %rbx, %rdi
|
|
2903
|
+
testq %rdx, %rcx
|
|
2904
|
+
cmoveq %rbp, %r8
|
|
2905
|
+
cmoveq %rbp, %rdi
|
|
2906
|
+
sarq $1, %rcx
|
|
2907
|
+
xorq %r8, %rdi
|
|
2908
|
+
xorq %r8, %rsi
|
|
2909
|
+
btq $0x3f, %r8
|
|
2910
|
+
cmovbq %rcx, %rbx
|
|
2911
|
+
movq %rax, %r8
|
|
2912
|
+
subq %rax, %rsi
|
|
2913
|
+
leaq (%rcx,%rdi), %rcx
|
|
2914
|
+
cmovs %rbp, %r8
|
|
2915
|
+
movq %rbx, %rdi
|
|
2916
|
+
testq %rdx, %rcx
|
|
2917
|
+
cmoveq %rbp, %r8
|
|
2918
|
+
cmoveq %rbp, %rdi
|
|
2919
|
+
sarq $1, %rcx
|
|
2920
|
+
xorq %r8, %rdi
|
|
2921
|
+
xorq %r8, %rsi
|
|
2922
|
+
btq $0x3f, %r8
|
|
2923
|
+
cmovbq %rcx, %rbx
|
|
2924
|
+
movq %rax, %r8
|
|
2925
|
+
subq %rax, %rsi
|
|
2926
|
+
leaq (%rcx,%rdi), %rcx
|
|
2927
|
+
cmovs %rbp, %r8
|
|
2928
|
+
movq %rbx, %rdi
|
|
2929
|
+
testq %rdx, %rcx
|
|
2930
|
+
cmoveq %rbp, %r8
|
|
2931
|
+
cmoveq %rbp, %rdi
|
|
2932
|
+
sarq $1, %rcx
|
|
2933
|
+
xorq %r8, %rdi
|
|
2934
|
+
xorq %r8, %rsi
|
|
2935
|
+
btq $0x3f, %r8
|
|
2936
|
+
cmovbq %rcx, %rbx
|
|
2937
|
+
movq %rax, %r8
|
|
2938
|
+
subq %rax, %rsi
|
|
2939
|
+
leaq (%rcx,%rdi), %rcx
|
|
2940
|
+
sarq $1, %rcx
|
|
2941
|
+
movl $0x100000, %eax
|
|
2942
|
+
leaq (%rbx,%rax), %r8
|
|
2943
|
+
leaq (%rcx,%rax), %r10
|
|
2944
|
+
shlq $0x16, %r8
|
|
2945
|
+
shlq $0x16, %r10
|
|
2946
|
+
sarq $0x2b, %r8
|
|
2947
|
+
sarq $0x2b, %r10
|
|
2948
|
+
movabsq $0x20000100000, %rax
|
|
2949
|
+
leaq (%rbx,%rax), %r15
|
|
2950
|
+
leaq (%rcx,%rax), %r11
|
|
2951
|
+
sarq $0x2a, %r15
|
|
2952
|
+
sarq $0x2a, %r11
|
|
2953
|
+
movq %r13, %rbx
|
|
2954
|
+
movq %r12, %rcx
|
|
2955
|
+
imulq %r8, %r12
|
|
2956
|
+
imulq %r15, %rbx
|
|
2957
|
+
addq %rbx, %r12
|
|
2958
|
+
imulq %r11, %r13
|
|
2959
|
+
imulq %r10, %rcx
|
|
2960
|
+
addq %rcx, %r13
|
|
2961
|
+
sarq $0x14, %r12
|
|
2962
|
+
sarq $0x14, %r13
|
|
2963
|
+
movq %r12, %rbx
|
|
2964
|
+
andq $0xfffff, %rbx
|
|
2965
|
+
movabsq $0xfffffe0000000000, %rax
|
|
2966
|
+
orq %rax, %rbx
|
|
2967
|
+
movq %r13, %rcx
|
|
2968
|
+
andq $0xfffff, %rcx
|
|
2969
|
+
movabsq $0xc000000000000000, %rax
|
|
2970
|
+
orq %rax, %rcx
|
|
2971
|
+
movq 0xa0(%rsp), %rax
|
|
2972
|
+
imulq %r8, %rax
|
|
2973
|
+
movq 0xb0(%rsp), %rdx
|
|
2974
|
+
imulq %r15, %rdx
|
|
2975
|
+
imulq 0xa8(%rsp), %r8
|
|
2976
|
+
imulq 0xb8(%rsp), %r15
|
|
2977
|
+
addq %r8, %r15
|
|
2978
|
+
leaq (%rax,%rdx), %r9
|
|
2979
|
+
movq 0xa0(%rsp), %rax
|
|
2980
|
+
imulq %r10, %rax
|
|
2981
|
+
movq 0xb0(%rsp), %rdx
|
|
2982
|
+
imulq %r11, %rdx
|
|
2983
|
+
imulq 0xa8(%rsp), %r10
|
|
2984
|
+
imulq 0xb8(%rsp), %r11
|
|
2985
|
+
addq %r10, %r11
|
|
2986
|
+
leaq (%rax,%rdx), %r13
|
|
2987
|
+
movq $0xfffffffffffffffe, %rax
|
|
2988
|
+
movl $0x2, %edx
|
|
2989
|
+
movq %rbx, %rdi
|
|
2990
|
+
movq %rax, %r8
|
|
2991
|
+
testq %rsi, %rsi
|
|
2992
|
+
cmovs %rbp, %r8
|
|
2993
|
+
testq $0x1, %rcx
|
|
2994
|
+
cmoveq %rbp, %r8
|
|
2995
|
+
cmoveq %rbp, %rdi
|
|
2996
|
+
xorq %r8, %rdi
|
|
2997
|
+
xorq %r8, %rsi
|
|
2998
|
+
btq $0x3f, %r8
|
|
2999
|
+
cmovbq %rcx, %rbx
|
|
3000
|
+
movq %rax, %r8
|
|
3001
|
+
subq %rax, %rsi
|
|
3002
|
+
leaq (%rcx,%rdi), %rcx
|
|
3003
|
+
cmovs %rbp, %r8
|
|
3004
|
+
movq %rbx, %rdi
|
|
3005
|
+
testq %rdx, %rcx
|
|
3006
|
+
cmoveq %rbp, %r8
|
|
3007
|
+
cmoveq %rbp, %rdi
|
|
3008
|
+
sarq $1, %rcx
|
|
3009
|
+
xorq %r8, %rdi
|
|
3010
|
+
xorq %r8, %rsi
|
|
3011
|
+
btq $0x3f, %r8
|
|
3012
|
+
cmovbq %rcx, %rbx
|
|
3013
|
+
movq %rax, %r8
|
|
3014
|
+
subq %rax, %rsi
|
|
3015
|
+
leaq (%rcx,%rdi), %rcx
|
|
3016
|
+
cmovs %rbp, %r8
|
|
3017
|
+
movq %rbx, %rdi
|
|
3018
|
+
testq %rdx, %rcx
|
|
3019
|
+
cmoveq %rbp, %r8
|
|
3020
|
+
cmoveq %rbp, %rdi
|
|
3021
|
+
sarq $1, %rcx
|
|
3022
|
+
xorq %r8, %rdi
|
|
3023
|
+
xorq %r8, %rsi
|
|
3024
|
+
btq $0x3f, %r8
|
|
3025
|
+
cmovbq %rcx, %rbx
|
|
3026
|
+
movq %rax, %r8
|
|
3027
|
+
subq %rax, %rsi
|
|
3028
|
+
leaq (%rcx,%rdi), %rcx
|
|
3029
|
+
cmovs %rbp, %r8
|
|
3030
|
+
movq %rbx, %rdi
|
|
3031
|
+
testq %rdx, %rcx
|
|
3032
|
+
cmoveq %rbp, %r8
|
|
3033
|
+
cmoveq %rbp, %rdi
|
|
3034
|
+
sarq $1, %rcx
|
|
3035
|
+
xorq %r8, %rdi
|
|
3036
|
+
xorq %r8, %rsi
|
|
3037
|
+
btq $0x3f, %r8
|
|
3038
|
+
cmovbq %rcx, %rbx
|
|
3039
|
+
movq %rax, %r8
|
|
3040
|
+
subq %rax, %rsi
|
|
3041
|
+
leaq (%rcx,%rdi), %rcx
|
|
3042
|
+
cmovs %rbp, %r8
|
|
3043
|
+
movq %rbx, %rdi
|
|
3044
|
+
testq %rdx, %rcx
|
|
3045
|
+
cmoveq %rbp, %r8
|
|
3046
|
+
cmoveq %rbp, %rdi
|
|
3047
|
+
sarq $1, %rcx
|
|
3048
|
+
xorq %r8, %rdi
|
|
3049
|
+
xorq %r8, %rsi
|
|
3050
|
+
btq $0x3f, %r8
|
|
3051
|
+
cmovbq %rcx, %rbx
|
|
3052
|
+
movq %rax, %r8
|
|
3053
|
+
subq %rax, %rsi
|
|
3054
|
+
leaq (%rcx,%rdi), %rcx
|
|
3055
|
+
cmovs %rbp, %r8
|
|
3056
|
+
movq %rbx, %rdi
|
|
3057
|
+
testq %rdx, %rcx
|
|
3058
|
+
cmoveq %rbp, %r8
|
|
3059
|
+
cmoveq %rbp, %rdi
|
|
3060
|
+
sarq $1, %rcx
|
|
3061
|
+
xorq %r8, %rdi
|
|
3062
|
+
xorq %r8, %rsi
|
|
3063
|
+
btq $0x3f, %r8
|
|
3064
|
+
cmovbq %rcx, %rbx
|
|
3065
|
+
movq %rax, %r8
|
|
3066
|
+
subq %rax, %rsi
|
|
3067
|
+
leaq (%rcx,%rdi), %rcx
|
|
3068
|
+
cmovs %rbp, %r8
|
|
3069
|
+
movq %rbx, %rdi
|
|
3070
|
+
testq %rdx, %rcx
|
|
3071
|
+
cmoveq %rbp, %r8
|
|
3072
|
+
cmoveq %rbp, %rdi
|
|
3073
|
+
sarq $1, %rcx
|
|
3074
|
+
xorq %r8, %rdi
|
|
3075
|
+
xorq %r8, %rsi
|
|
3076
|
+
btq $0x3f, %r8
|
|
3077
|
+
cmovbq %rcx, %rbx
|
|
3078
|
+
movq %rax, %r8
|
|
3079
|
+
subq %rax, %rsi
|
|
3080
|
+
leaq (%rcx,%rdi), %rcx
|
|
3081
|
+
cmovs %rbp, %r8
|
|
3082
|
+
movq %rbx, %rdi
|
|
3083
|
+
testq %rdx, %rcx
|
|
3084
|
+
cmoveq %rbp, %r8
|
|
3085
|
+
cmoveq %rbp, %rdi
|
|
3086
|
+
sarq $1, %rcx
|
|
3087
|
+
xorq %r8, %rdi
|
|
3088
|
+
xorq %r8, %rsi
|
|
3089
|
+
btq $0x3f, %r8
|
|
3090
|
+
cmovbq %rcx, %rbx
|
|
3091
|
+
movq %rax, %r8
|
|
3092
|
+
subq %rax, %rsi
|
|
3093
|
+
leaq (%rcx,%rdi), %rcx
|
|
3094
|
+
cmovs %rbp, %r8
|
|
3095
|
+
movq %rbx, %rdi
|
|
3096
|
+
testq %rdx, %rcx
|
|
3097
|
+
cmoveq %rbp, %r8
|
|
3098
|
+
cmoveq %rbp, %rdi
|
|
3099
|
+
sarq $1, %rcx
|
|
3100
|
+
xorq %r8, %rdi
|
|
3101
|
+
xorq %r8, %rsi
|
|
3102
|
+
btq $0x3f, %r8
|
|
3103
|
+
cmovbq %rcx, %rbx
|
|
3104
|
+
movq %rax, %r8
|
|
3105
|
+
subq %rax, %rsi
|
|
3106
|
+
leaq (%rcx,%rdi), %rcx
|
|
3107
|
+
cmovs %rbp, %r8
|
|
3108
|
+
movq %rbx, %rdi
|
|
3109
|
+
testq %rdx, %rcx
|
|
3110
|
+
cmoveq %rbp, %r8
|
|
3111
|
+
cmoveq %rbp, %rdi
|
|
3112
|
+
sarq $1, %rcx
|
|
3113
|
+
xorq %r8, %rdi
|
|
3114
|
+
xorq %r8, %rsi
|
|
3115
|
+
btq $0x3f, %r8
|
|
3116
|
+
cmovbq %rcx, %rbx
|
|
3117
|
+
movq %rax, %r8
|
|
3118
|
+
subq %rax, %rsi
|
|
3119
|
+
leaq (%rcx,%rdi), %rcx
|
|
3120
|
+
cmovs %rbp, %r8
|
|
3121
|
+
movq %rbx, %rdi
|
|
3122
|
+
testq %rdx, %rcx
|
|
3123
|
+
cmoveq %rbp, %r8
|
|
3124
|
+
cmoveq %rbp, %rdi
|
|
3125
|
+
sarq $1, %rcx
|
|
3126
|
+
xorq %r8, %rdi
|
|
3127
|
+
xorq %r8, %rsi
|
|
3128
|
+
btq $0x3f, %r8
|
|
3129
|
+
cmovbq %rcx, %rbx
|
|
3130
|
+
movq %rax, %r8
|
|
3131
|
+
subq %rax, %rsi
|
|
3132
|
+
leaq (%rcx,%rdi), %rcx
|
|
3133
|
+
cmovs %rbp, %r8
|
|
3134
|
+
movq %rbx, %rdi
|
|
3135
|
+
testq %rdx, %rcx
|
|
3136
|
+
cmoveq %rbp, %r8
|
|
3137
|
+
cmoveq %rbp, %rdi
|
|
3138
|
+
sarq $1, %rcx
|
|
3139
|
+
xorq %r8, %rdi
|
|
3140
|
+
xorq %r8, %rsi
|
|
3141
|
+
btq $0x3f, %r8
|
|
3142
|
+
cmovbq %rcx, %rbx
|
|
3143
|
+
movq %rax, %r8
|
|
3144
|
+
subq %rax, %rsi
|
|
3145
|
+
leaq (%rcx,%rdi), %rcx
|
|
3146
|
+
cmovs %rbp, %r8
|
|
3147
|
+
movq %rbx, %rdi
|
|
3148
|
+
testq %rdx, %rcx
|
|
3149
|
+
cmoveq %rbp, %r8
|
|
3150
|
+
cmoveq %rbp, %rdi
|
|
3151
|
+
sarq $1, %rcx
|
|
3152
|
+
xorq %r8, %rdi
|
|
3153
|
+
xorq %r8, %rsi
|
|
3154
|
+
btq $0x3f, %r8
|
|
3155
|
+
cmovbq %rcx, %rbx
|
|
3156
|
+
movq %rax, %r8
|
|
3157
|
+
subq %rax, %rsi
|
|
3158
|
+
leaq (%rcx,%rdi), %rcx
|
|
3159
|
+
cmovs %rbp, %r8
|
|
3160
|
+
movq %rbx, %rdi
|
|
3161
|
+
testq %rdx, %rcx
|
|
3162
|
+
cmoveq %rbp, %r8
|
|
3163
|
+
cmoveq %rbp, %rdi
|
|
3164
|
+
sarq $1, %rcx
|
|
3165
|
+
xorq %r8, %rdi
|
|
3166
|
+
xorq %r8, %rsi
|
|
3167
|
+
btq $0x3f, %r8
|
|
3168
|
+
cmovbq %rcx, %rbx
|
|
3169
|
+
movq %rax, %r8
|
|
3170
|
+
subq %rax, %rsi
|
|
3171
|
+
leaq (%rcx,%rdi), %rcx
|
|
3172
|
+
cmovs %rbp, %r8
|
|
3173
|
+
movq %rbx, %rdi
|
|
3174
|
+
testq %rdx, %rcx
|
|
3175
|
+
cmoveq %rbp, %r8
|
|
3176
|
+
cmoveq %rbp, %rdi
|
|
3177
|
+
sarq $1, %rcx
|
|
3178
|
+
xorq %r8, %rdi
|
|
3179
|
+
xorq %r8, %rsi
|
|
3180
|
+
btq $0x3f, %r8
|
|
3181
|
+
cmovbq %rcx, %rbx
|
|
3182
|
+
movq %rax, %r8
|
|
3183
|
+
subq %rax, %rsi
|
|
3184
|
+
leaq (%rcx,%rdi), %rcx
|
|
3185
|
+
cmovs %rbp, %r8
|
|
3186
|
+
movq %rbx, %rdi
|
|
3187
|
+
testq %rdx, %rcx
|
|
3188
|
+
cmoveq %rbp, %r8
|
|
3189
|
+
cmoveq %rbp, %rdi
|
|
3190
|
+
sarq $1, %rcx
|
|
3191
|
+
xorq %r8, %rdi
|
|
3192
|
+
xorq %r8, %rsi
|
|
3193
|
+
btq $0x3f, %r8
|
|
3194
|
+
cmovbq %rcx, %rbx
|
|
3195
|
+
movq %rax, %r8
|
|
3196
|
+
subq %rax, %rsi
|
|
3197
|
+
leaq (%rcx,%rdi), %rcx
|
|
3198
|
+
cmovs %rbp, %r8
|
|
3199
|
+
movq %rbx, %rdi
|
|
3200
|
+
testq %rdx, %rcx
|
|
3201
|
+
cmoveq %rbp, %r8
|
|
3202
|
+
cmoveq %rbp, %rdi
|
|
3203
|
+
sarq $1, %rcx
|
|
3204
|
+
xorq %r8, %rdi
|
|
3205
|
+
xorq %r8, %rsi
|
|
3206
|
+
btq $0x3f, %r8
|
|
3207
|
+
cmovbq %rcx, %rbx
|
|
3208
|
+
movq %rax, %r8
|
|
3209
|
+
subq %rax, %rsi
|
|
3210
|
+
leaq (%rcx,%rdi), %rcx
|
|
3211
|
+
cmovs %rbp, %r8
|
|
3212
|
+
movq %rbx, %rdi
|
|
3213
|
+
testq %rdx, %rcx
|
|
3214
|
+
cmoveq %rbp, %r8
|
|
3215
|
+
cmoveq %rbp, %rdi
|
|
3216
|
+
sarq $1, %rcx
|
|
3217
|
+
xorq %r8, %rdi
|
|
3218
|
+
xorq %r8, %rsi
|
|
3219
|
+
btq $0x3f, %r8
|
|
3220
|
+
cmovbq %rcx, %rbx
|
|
3221
|
+
movq %rax, %r8
|
|
3222
|
+
subq %rax, %rsi
|
|
3223
|
+
leaq (%rcx,%rdi), %rcx
|
|
3224
|
+
cmovs %rbp, %r8
|
|
3225
|
+
movq %rbx, %rdi
|
|
3226
|
+
testq %rdx, %rcx
|
|
3227
|
+
cmoveq %rbp, %r8
|
|
3228
|
+
cmoveq %rbp, %rdi
|
|
3229
|
+
sarq $1, %rcx
|
|
3230
|
+
xorq %r8, %rdi
|
|
3231
|
+
xorq %r8, %rsi
|
|
3232
|
+
btq $0x3f, %r8
|
|
3233
|
+
cmovbq %rcx, %rbx
|
|
3234
|
+
movq %rax, %r8
|
|
3235
|
+
subq %rax, %rsi
|
|
3236
|
+
leaq (%rcx,%rdi), %rcx
|
|
3237
|
+
sarq $1, %rcx
|
|
3238
|
+
movl $0x100000, %eax
|
|
3239
|
+
leaq (%rbx,%rax), %r8
|
|
3240
|
+
leaq (%rcx,%rax), %r12
|
|
3241
|
+
shlq $0x15, %r8
|
|
3242
|
+
shlq $0x15, %r12
|
|
3243
|
+
sarq $0x2b, %r8
|
|
3244
|
+
sarq $0x2b, %r12
|
|
3245
|
+
movabsq $0x20000100000, %rax
|
|
3246
|
+
leaq (%rbx,%rax), %r10
|
|
3247
|
+
leaq (%rcx,%rax), %r14
|
|
3248
|
+
sarq $0x2b, %r10
|
|
3249
|
+
sarq $0x2b, %r14
|
|
3250
|
+
movq %r9, %rax
|
|
3251
|
+
imulq %r8, %rax
|
|
3252
|
+
movq %r13, %rdx
|
|
3253
|
+
imulq %r10, %rdx
|
|
3254
|
+
imulq %r15, %r8
|
|
3255
|
+
imulq %r11, %r10
|
|
3256
|
+
addq %r8, %r10
|
|
3257
|
+
leaq (%rax,%rdx), %r8
|
|
3258
|
+
movq %r9, %rax
|
|
3259
|
+
imulq %r12, %rax
|
|
3260
|
+
movq %r13, %rdx
|
|
3261
|
+
imulq %r14, %rdx
|
|
3262
|
+
imulq %r15, %r12
|
|
3263
|
+
imulq %r11, %r14
|
|
3264
|
+
addq %r12, %r14
|
|
3265
|
+
leaq (%rax,%rdx), %r12
|
|
3266
|
+
movq %rsi, 0x98(%rsp)
|
|
3267
|
+
decq 0x90(%rsp)
|
|
3268
|
+
jne edwards25519_scalarmuldouble_inverseloop
|
|
3269
|
+
movq (%rsp), %rax
|
|
3270
|
+
movq 0x20(%rsp), %rcx
|
|
3271
|
+
imulq %r8, %rax
|
|
3272
|
+
imulq %r10, %rcx
|
|
3273
|
+
addq %rcx, %rax
|
|
3274
|
+
sarq $0x3f, %rax
|
|
3275
|
+
movq %r8, %r9
|
|
3276
|
+
sarq $0x3f, %r9
|
|
3277
|
+
xorq %r9, %r8
|
|
3278
|
+
subq %r9, %r8
|
|
3279
|
+
xorq %rax, %r9
|
|
3280
|
+
movq %r10, %r11
|
|
3281
|
+
sarq $0x3f, %r11
|
|
3282
|
+
xorq %r11, %r10
|
|
3283
|
+
subq %r11, %r10
|
|
3284
|
+
xorq %rax, %r11
|
|
3285
|
+
movq %r12, %r13
|
|
3286
|
+
sarq $0x3f, %r13
|
|
3287
|
+
xorq %r13, %r12
|
|
3288
|
+
subq %r13, %r12
|
|
3289
|
+
xorq %rax, %r13
|
|
3290
|
+
movq %r14, %r15
|
|
3291
|
+
sarq $0x3f, %r15
|
|
3292
|
+
xorq %r15, %r14
|
|
3293
|
+
subq %r15, %r14
|
|
3294
|
+
xorq %rax, %r15
|
|
3295
|
+
movq %r8, %rax
|
|
3296
|
+
andq %r9, %rax
|
|
3297
|
+
movq %r10, %r12
|
|
3298
|
+
andq %r11, %r12
|
|
3299
|
+
addq %rax, %r12
|
|
3300
|
+
xorl %r13d, %r13d
|
|
3301
|
+
movq 0x40(%rsp), %rax
|
|
3302
|
+
xorq %r9, %rax
|
|
3303
|
+
mulq %r8
|
|
3304
|
+
addq %rax, %r12
|
|
3305
|
+
adcq %rdx, %r13
|
|
3306
|
+
movq 0x60(%rsp), %rax
|
|
3307
|
+
xorq %r11, %rax
|
|
3308
|
+
mulq %r10
|
|
3309
|
+
addq %rax, %r12
|
|
3310
|
+
adcq %rdx, %r13
|
|
3311
|
+
xorl %r14d, %r14d
|
|
3312
|
+
movq 0x48(%rsp), %rax
|
|
3313
|
+
xorq %r9, %rax
|
|
3314
|
+
mulq %r8
|
|
3315
|
+
addq %rax, %r13
|
|
3316
|
+
adcq %rdx, %r14
|
|
3317
|
+
movq 0x68(%rsp), %rax
|
|
3318
|
+
xorq %r11, %rax
|
|
3319
|
+
mulq %r10
|
|
3320
|
+
addq %rax, %r13
|
|
3321
|
+
adcq %rdx, %r14
|
|
3322
|
+
xorl %r15d, %r15d
|
|
3323
|
+
movq 0x50(%rsp), %rax
|
|
3324
|
+
xorq %r9, %rax
|
|
3325
|
+
mulq %r8
|
|
3326
|
+
addq %rax, %r14
|
|
3327
|
+
adcq %rdx, %r15
|
|
3328
|
+
movq 0x70(%rsp), %rax
|
|
3329
|
+
xorq %r11, %rax
|
|
3330
|
+
mulq %r10
|
|
3331
|
+
addq %rax, %r14
|
|
3332
|
+
adcq %rdx, %r15
|
|
3333
|
+
movq 0x58(%rsp), %rax
|
|
3334
|
+
xorq %r9, %rax
|
|
3335
|
+
andq %r8, %r9
|
|
3336
|
+
negq %r9
|
|
3337
|
+
mulq %r8
|
|
3338
|
+
addq %rax, %r15
|
|
3339
|
+
adcq %rdx, %r9
|
|
3340
|
+
movq 0x78(%rsp), %rax
|
|
3341
|
+
xorq %r11, %rax
|
|
3342
|
+
movq %r11, %rdx
|
|
3343
|
+
andq %r10, %rdx
|
|
3344
|
+
subq %rdx, %r9
|
|
3345
|
+
mulq %r10
|
|
3346
|
+
addq %rax, %r15
|
|
3347
|
+
adcq %rdx, %r9
|
|
3348
|
+
movq %r9, %rax
|
|
3349
|
+
shldq $0x1, %r15, %rax
|
|
3350
|
+
sarq $0x3f, %r9
|
|
3351
|
+
movl $0x13, %ebx
|
|
3352
|
+
leaq 0x1(%rax,%r9,1), %rax
|
|
3353
|
+
imulq %rbx
|
|
3354
|
+
xorl %ebp, %ebp
|
|
3355
|
+
addq %rax, %r12
|
|
3356
|
+
adcq %rdx, %r13
|
|
3357
|
+
adcq %r9, %r14
|
|
3358
|
+
adcq %r9, %r15
|
|
3359
|
+
shlq $0x3f, %rax
|
|
3360
|
+
addq %rax, %r15
|
|
3361
|
+
cmovns %rbp, %rbx
|
|
3362
|
+
subq %rbx, %r12
|
|
3363
|
+
sbbq %rbp, %r13
|
|
3364
|
+
sbbq %rbp, %r14
|
|
3365
|
+
sbbq %rbp, %r15
|
|
3366
|
+
btr $0x3f, %r15
|
|
3367
|
+
movq 0xc0(%rsp), %rdi
|
|
3368
|
+
movq %r12, (%rdi)
|
|
3369
|
+
movq %r13, 0x8(%rdi)
|
|
3370
|
+
movq %r14, 0x10(%rdi)
|
|
3371
|
+
movq %r15, 0x18(%rdi)
|
|
3372
|
+
|
|
3373
|
+
// Store result
|
|
3374
|
+
|
|
3375
|
+
movq res, %rdi
|
|
3376
|
+
leaq ACC(%rsp), %rsi
|
|
3377
|
+
leaq TAB(%rsp), %rbp
|
|
3378
|
+
mul_p25519(x_0,x_1,x_2)
|
|
3379
|
+
|
|
3380
|
+
movq res, %rdi
|
|
3381
|
+
addq $32, %rdi
|
|
3382
|
+
leaq ACC+32(%rsp), %rsi
|
|
3383
|
+
leaq TAB(%rsp), %rbp
|
|
3384
|
+
mul_p25519(x_0,x_1,x_2)
|
|
3385
|
+
|
|
3386
|
+
// Restore stack and registers
|
|
3387
|
+
|
|
3388
|
+
addq $NSPACE, %rsp
|
|
3389
|
+
|
|
3390
|
+
popq %r15
|
|
3391
|
+
popq %r14
|
|
3392
|
+
popq %r13
|
|
3393
|
+
popq %r12
|
|
3394
|
+
popq %rbp
|
|
3395
|
+
popq %rbx
|
|
3396
|
+
ret
|
|
3397
|
+
|
|
3398
|
+
// ****************************************************************************
|
|
3399
|
+
// Localized versions of subroutines.
|
|
3400
|
+
// These are close to the standalone functions "edwards25519_epdouble" etc.,
|
|
3401
|
+
// but are only maintaining reduction modulo 2^256 - 38, not 2^255 - 19.
|
|
3402
|
+
// ****************************************************************************
|
|
3403
|
+
|
|
3404
|
+
edwards25519_scalarmuldouble_epdouble:
|
|
3405
|
+
sub $(5*NUMSIZE), %rsp
|
|
3406
|
+
add_twice4(t0,x_1,y_1)
|
|
3407
|
+
sqr_4(t1,z_1)
|
|
3408
|
+
sqr_4(t2,x_1)
|
|
3409
|
+
sqr_4(t3,y_1)
|
|
3410
|
+
double_twice4(t1,t1)
|
|
3411
|
+
sqr_4(t0,t0)
|
|
3412
|
+
add_twice4(t4,t2,t3)
|
|
3413
|
+
sub_twice4(t2,t2,t3)
|
|
3414
|
+
add_twice4(t3,t1,t2)
|
|
3415
|
+
sub_twice4(t1,t4,t0)
|
|
3416
|
+
mul_4(y_0,t2,t4)
|
|
3417
|
+
mul_4(z_0,t3,t2)
|
|
3418
|
+
mul_4(w_0,t1,t4)
|
|
3419
|
+
mul_4(x_0,t1,t3)
|
|
3420
|
+
add $(5*NUMSIZE), %rsp
|
|
3421
|
+
ret
|
|
3422
|
+
|
|
3423
|
+
edwards25519_scalarmuldouble_pdouble:
|
|
3424
|
+
sub $(5*NUMSIZE), %rsp
|
|
3425
|
+
add_twice4(t0,x_1,y_1)
|
|
3426
|
+
sqr_4(t1,z_1)
|
|
3427
|
+
sqr_4(t2,x_1)
|
|
3428
|
+
sqr_4(t3,y_1)
|
|
3429
|
+
double_twice4(t1,t1)
|
|
3430
|
+
sqr_4(t0,t0)
|
|
3431
|
+
add_twice4(t4,t2,t3)
|
|
3432
|
+
sub_twice4(t2,t2,t3)
|
|
3433
|
+
add_twice4(t3,t1,t2)
|
|
3434
|
+
sub_twice4(t1,t4,t0)
|
|
3435
|
+
mul_4(y_0,t2,t4)
|
|
3436
|
+
mul_4(z_0,t3,t2)
|
|
3437
|
+
mul_4(x_0,t1,t3)
|
|
3438
|
+
add $(5*NUMSIZE), %rsp
|
|
3439
|
+
ret
|
|
3440
|
+
|
|
3441
|
+
edwards25519_scalarmuldouble_epadd:
|
|
3442
|
+
sub $(6*NUMSIZE), %rsp
|
|
3443
|
+
mul_4(t0,w_1,w_2)
|
|
3444
|
+
sub_twice4(t1,y_1,x_1)
|
|
3445
|
+
sub_twice4(t2,y_2,x_2)
|
|
3446
|
+
add_twice4(t3,y_1,x_1)
|
|
3447
|
+
add_twice4(t4,y_2,x_2)
|
|
3448
|
+
double_twice4(t5,z_2)
|
|
3449
|
+
mul_4(t1,t1,t2)
|
|
3450
|
+
mul_4(t3,t3,t4)
|
|
3451
|
+
load_k25519(t2)
|
|
3452
|
+
mul_4(t2,t2,t0)
|
|
3453
|
+
mul_4(t4,z_1,t5)
|
|
3454
|
+
sub_twice4(t0,t3,t1)
|
|
3455
|
+
add_twice4(t5,t3,t1)
|
|
3456
|
+
sub_twice4(t1,t4,t2)
|
|
3457
|
+
add_twice4(t3,t4,t2)
|
|
3458
|
+
mul_4(w_0,t0,t5)
|
|
3459
|
+
mul_4(x_0,t0,t1)
|
|
3460
|
+
mul_4(y_0,t3,t5)
|
|
3461
|
+
mul_4(z_0,t1,t3)
|
|
3462
|
+
add $(6*NUMSIZE), %rsp
|
|
3463
|
+
ret
|
|
3464
|
+
|
|
3465
|
+
edwards25519_scalarmuldouble_pepadd:
|
|
3466
|
+
sub $(6*NUMSIZE), %rsp
|
|
3467
|
+
double_twice4(t0,z_1);
|
|
3468
|
+
sub_twice4(t1,y_1,x_1);
|
|
3469
|
+
add_twice4(t2,y_1,x_1);
|
|
3470
|
+
mul_4(t3,w_1,z_2);
|
|
3471
|
+
mul_4(t1,t1,x_2);
|
|
3472
|
+
mul_4(t2,t2,y_2);
|
|
3473
|
+
sub_twice4(t4,t0,t3);
|
|
3474
|
+
add_twice4(t0,t0,t3);
|
|
3475
|
+
sub_twice4(t5,t2,t1);
|
|
3476
|
+
add_twice4(t1,t2,t1);
|
|
3477
|
+
mul_4(z_0,t4,t0);
|
|
3478
|
+
mul_4(x_0,t5,t4);
|
|
3479
|
+
mul_4(y_0,t0,t1);
|
|
3480
|
+
mul_4(w_0,t5,t1);
|
|
3481
|
+
add $(6*NUMSIZE), %rsp
|
|
3482
|
+
ret
|
|
3483
|
+
|
|
3484
|
+
// ****************************************************************************
|
|
3485
|
+
// The precomputed data (all read-only). This is currently part of the same
|
|
3486
|
+
// text section, which gives position-independent code with simple PC-relative
|
|
3487
|
+
// addressing. However it could be put in a separate section via something like
|
|
3488
|
+
//
|
|
3489
|
+
// .section .rodata
|
|
3490
|
+
// ****************************************************************************
|
|
3491
|
+
|
|
3492
|
+
// Precomputed table of multiples of generator for edwards25519
|
|
3493
|
+
// all in precomputed extended-projective (y-x,x+y,2*d*x*y) triples.
|
|
3494
|
+
|
|
3495
|
+
edwards25519_scalarmuldouble_table:
|
|
3496
|
+
|
|
3497
|
+
// 1 * G
|
|
3498
|
+
|
|
3499
|
+
.quad 0x9d103905d740913e
|
|
3500
|
+
.quad 0xfd399f05d140beb3
|
|
3501
|
+
.quad 0xa5c18434688f8a09
|
|
3502
|
+
.quad 0x44fd2f9298f81267
|
|
3503
|
+
.quad 0x2fbc93c6f58c3b85
|
|
3504
|
+
.quad 0xcf932dc6fb8c0e19
|
|
3505
|
+
.quad 0x270b4898643d42c2
|
|
3506
|
+
.quad 0x07cf9d3a33d4ba65
|
|
3507
|
+
.quad 0xabc91205877aaa68
|
|
3508
|
+
.quad 0x26d9e823ccaac49e
|
|
3509
|
+
.quad 0x5a1b7dcbdd43598c
|
|
3510
|
+
.quad 0x6f117b689f0c65a8
|
|
3511
|
+
|
|
3512
|
+
// 2 * G
|
|
3513
|
+
|
|
3514
|
+
.quad 0x8a99a56042b4d5a8
|
|
3515
|
+
.quad 0x8f2b810c4e60acf6
|
|
3516
|
+
.quad 0xe09e236bb16e37aa
|
|
3517
|
+
.quad 0x6bb595a669c92555
|
|
3518
|
+
.quad 0x9224e7fc933c71d7
|
|
3519
|
+
.quad 0x9f469d967a0ff5b5
|
|
3520
|
+
.quad 0x5aa69a65e1d60702
|
|
3521
|
+
.quad 0x590c063fa87d2e2e
|
|
3522
|
+
.quad 0x43faa8b3a59b7a5f
|
|
3523
|
+
.quad 0x36c16bdd5d9acf78
|
|
3524
|
+
.quad 0x500fa0840b3d6a31
|
|
3525
|
+
.quad 0x701af5b13ea50b73
|
|
3526
|
+
|
|
3527
|
+
// 3 * G
|
|
3528
|
+
|
|
3529
|
+
.quad 0x56611fe8a4fcd265
|
|
3530
|
+
.quad 0x3bd353fde5c1ba7d
|
|
3531
|
+
.quad 0x8131f31a214bd6bd
|
|
3532
|
+
.quad 0x2ab91587555bda62
|
|
3533
|
+
.quad 0xaf25b0a84cee9730
|
|
3534
|
+
.quad 0x025a8430e8864b8a
|
|
3535
|
+
.quad 0xc11b50029f016732
|
|
3536
|
+
.quad 0x7a164e1b9a80f8f4
|
|
3537
|
+
.quad 0x14ae933f0dd0d889
|
|
3538
|
+
.quad 0x589423221c35da62
|
|
3539
|
+
.quad 0xd170e5458cf2db4c
|
|
3540
|
+
.quad 0x5a2826af12b9b4c6
|
|
3541
|
+
|
|
3542
|
+
// 4 * G
|
|
3543
|
+
|
|
3544
|
+
.quad 0x95fe050a056818bf
|
|
3545
|
+
.quad 0x327e89715660faa9
|
|
3546
|
+
.quad 0xc3e8e3cd06a05073
|
|
3547
|
+
.quad 0x27933f4c7445a49a
|
|
3548
|
+
.quad 0x287351b98efc099f
|
|
3549
|
+
.quad 0x6765c6f47dfd2538
|
|
3550
|
+
.quad 0xca348d3dfb0a9265
|
|
3551
|
+
.quad 0x680e910321e58727
|
|
3552
|
+
.quad 0x5a13fbe9c476ff09
|
|
3553
|
+
.quad 0x6e9e39457b5cc172
|
|
3554
|
+
.quad 0x5ddbdcf9102b4494
|
|
3555
|
+
.quad 0x7f9d0cbf63553e2b
|
|
3556
|
+
|
|
3557
|
+
// 5 * G
|
|
3558
|
+
|
|
3559
|
+
.quad 0x7f9182c3a447d6ba
|
|
3560
|
+
.quad 0xd50014d14b2729b7
|
|
3561
|
+
.quad 0xe33cf11cb864a087
|
|
3562
|
+
.quad 0x154a7e73eb1b55f3
|
|
3563
|
+
.quad 0xa212bc4408a5bb33
|
|
3564
|
+
.quad 0x8d5048c3c75eed02
|
|
3565
|
+
.quad 0xdd1beb0c5abfec44
|
|
3566
|
+
.quad 0x2945ccf146e206eb
|
|
3567
|
+
.quad 0xbcbbdbf1812a8285
|
|
3568
|
+
.quad 0x270e0807d0bdd1fc
|
|
3569
|
+
.quad 0xb41b670b1bbda72d
|
|
3570
|
+
.quad 0x43aabe696b3bb69a
|
|
3571
|
+
|
|
3572
|
+
// 6 * G
|
|
3573
|
+
|
|
3574
|
+
.quad 0x499806b67b7d8ca4
|
|
3575
|
+
.quad 0x575be28427d22739
|
|
3576
|
+
.quad 0xbb085ce7204553b9
|
|
3577
|
+
.quad 0x38b64c41ae417884
|
|
3578
|
+
.quad 0x3a0ceeeb77157131
|
|
3579
|
+
.quad 0x9b27158900c8af88
|
|
3580
|
+
.quad 0x8065b668da59a736
|
|
3581
|
+
.quad 0x51e57bb6a2cc38bd
|
|
3582
|
+
.quad 0x85ac326702ea4b71
|
|
3583
|
+
.quad 0xbe70e00341a1bb01
|
|
3584
|
+
.quad 0x53e4a24b083bc144
|
|
3585
|
+
.quad 0x10b8e91a9f0d61e3
|
|
3586
|
+
|
|
3587
|
+
// 7 * G
|
|
3588
|
+
|
|
3589
|
+
.quad 0xba6f2c9aaa3221b1
|
|
3590
|
+
.quad 0x6ca021533bba23a7
|
|
3591
|
+
.quad 0x9dea764f92192c3a
|
|
3592
|
+
.quad 0x1d6edd5d2e5317e0
|
|
3593
|
+
.quad 0x6b1a5cd0944ea3bf
|
|
3594
|
+
.quad 0x7470353ab39dc0d2
|
|
3595
|
+
.quad 0x71b2528228542e49
|
|
3596
|
+
.quad 0x461bea69283c927e
|
|
3597
|
+
.quad 0xf1836dc801b8b3a2
|
|
3598
|
+
.quad 0xb3035f47053ea49a
|
|
3599
|
+
.quad 0x529c41ba5877adf3
|
|
3600
|
+
.quad 0x7a9fbb1c6a0f90a7
|
|
3601
|
+
|
|
3602
|
+
// 8 * G
|
|
3603
|
+
|
|
3604
|
+
.quad 0xe2a75dedf39234d9
|
|
3605
|
+
.quad 0x963d7680e1b558f9
|
|
3606
|
+
.quad 0x2c2741ac6e3c23fb
|
|
3607
|
+
.quad 0x3a9024a1320e01c3
|
|
3608
|
+
.quad 0x59b7596604dd3e8f
|
|
3609
|
+
.quad 0x6cb30377e288702c
|
|
3610
|
+
.quad 0xb1339c665ed9c323
|
|
3611
|
+
.quad 0x0915e76061bce52f
|
|
3612
|
+
.quad 0xe7c1f5d9c9a2911a
|
|
3613
|
+
.quad 0xb8a371788bcca7d7
|
|
3614
|
+
.quad 0x636412190eb62a32
|
|
3615
|
+
.quad 0x26907c5c2ecc4e95
|
|
3616
|
+
|
|
3617
|
+
#if defined(__linux__) && defined(__ELF__)
|
|
3618
|
+
.section .note.GNU-stack, "", %progbits
|
|
3619
|
+
#endif
|