commit fa8d5b5f83d66760c833d887e9ada0f300bcbfc7
parent 080d6d83136b1a7955e35f6f2bff74a0b3760cb7
Author: Armin Preiml <apreiml@strohwolke.at>
Date: Fri, 26 Nov 2021 13:34:05 +0100
implement AES-NI for x86_64
Signed-off-by: Armin Preiml <apreiml@strohwolke.at>
Signed-off-by: Drew DeVault <sir@cmpwn.com>
Diffstat:
5 files changed, 852 insertions(+), 5 deletions(-)
diff --git a/crypto/aes/+test/ni+x86_64.ha b/crypto/aes/+test/ni+x86_64.ha
@@ -0,0 +1,280 @@
+use bytes;
+use crypto::cipher;
+
+const zero_rk: [RKLEN_256]u8 = [0...];
+
+// taken from fips-197.pdf Section A.1
+@test fn ni_enc_key_expand_128() void = {
+ const key: [16]u8 = [
+ 0x2b, 0x7e, 0x15, 0x16, 0x28, 0xae, 0xd2, 0xa6,
+ 0xab, 0xf7, 0x15, 0x88, 0x09, 0xcf, 0x4f, 0x3c
+ ];
+
+ const expected_rounds: [_]u8 = [
+ 0x2b, 0x7e, 0x15, 0x16, 0x28, 0xae, 0xd2, 0xa6,
+ 0xab, 0xf7, 0x15, 0x88, 0x09, 0xcf, 0x4f, 0x3c,
+ 0xa0, 0xfa, 0xfe, 0x17, 0x88, 0x54, 0x2c, 0xb1,
+ 0x23, 0xa3, 0x39, 0x39, 0x2a, 0x6c, 0x76, 0x05,
+ 0xf2, 0xc2, 0x95, 0xf2, 0x7a, 0x96, 0xb9, 0x43,
+ 0x59, 0x35, 0x80, 0x7a, 0x73, 0x59, 0xf6, 0x7f,
+ 0x3d, 0x80, 0x47, 0x7d, 0x47, 0x16, 0xfe, 0x3e,
+ 0x1e, 0x23, 0x7e, 0x44, 0x6d, 0x7a, 0x88, 0x3b,
+ 0xef, 0x44, 0xa5, 0x41, 0xa8, 0x52, 0x5b, 0x7f,
+ 0xb6, 0x71, 0x25, 0x3b, 0xdb, 0x0b, 0xad, 0x00,
+ 0xd4, 0xd1, 0xc6, 0xf8, 0x7c, 0x83, 0x9d, 0x87,
+ 0xca, 0xf2, 0xb8, 0xbc, 0x11, 0xf9, 0x15, 0xbc,
+ 0x6d, 0x88, 0xa3, 0x7a, 0x11, 0x0b, 0x3e, 0xfd,
+ 0xdb, 0xf9, 0x86, 0x41, 0xca, 0x00, 0x93, 0xfd,
+ 0x4e, 0x54, 0xf7, 0x0e, 0x5f, 0x5f, 0xc9, 0xf3,
+ 0x84, 0xa6, 0x4f, 0xb2, 0x4e, 0xa6, 0xdc, 0x4f,
+ 0xea, 0xd2, 0x73, 0x21, 0xb5, 0x8d, 0xba, 0xd2,
+ 0x31, 0x2b, 0xf5, 0x60, 0x7f, 0x8d, 0x29, 0x2f,
+ 0xac, 0x77, 0x66, 0xf3, 0x19, 0xfa, 0xdc, 0x21,
+ 0x28, 0xd1, 0x29, 0x41, 0x57, 0x5c, 0x00, 0x6e,
+ 0xd0, 0x14, 0xf9, 0xa8, 0xc9, 0xee, 0x25, 0x89,
+ 0xe1, 0x3f, 0x0c, 0xc8, 0xb6, 0x63, 0x0c, 0xa6,
+ ];
+
+ let block = x86ni();
+ x86ni_init(&block, key[..]);
+
+ assert(block.rklen == 176);
+ assert(bytes::equal(expected_rounds[..], block.enc_rk[..RKLEN_128]));
+
+ cipher::finish(&block);
+ assert(bytes::equal(zero_rk[..], block.enc_rk[..]));
+ assert(bytes::equal(zero_rk[..], block.dec_rk[..]));
+};
+
+// taken from fips-197.pdf Section A.2
+@test fn ni_enc_key_expand_192() void = {
+ const key: [24]u8 = [
+ 0x8e, 0x73, 0xb0, 0xf7, 0xda, 0x0e, 0x64, 0x52,
+ 0xc8, 0x10, 0xf3, 0x2b, 0x80, 0x90, 0x79, 0xe5,
+ 0x62, 0xf8, 0xea, 0xd2, 0x52, 0x2c, 0x6b, 0x7b
+ ];
+
+ const expected_rounds: [_]u8 = [
+ 0x8e, 0x73, 0xb0, 0xf7, 0xda, 0x0e, 0x64, 0x52,
+ 0xc8, 0x10, 0xf3, 0x2b, 0x80, 0x90, 0x79, 0xe5,
+ 0x62, 0xf8, 0xea, 0xd2, 0x52, 0x2c, 0x6b, 0x7b,
+ 0xfe, 0x0c, 0x91, 0xf7, 0x24, 0x02, 0xf5, 0xa5,
+ 0xec, 0x12, 0x06, 0x8e, 0x6c, 0x82, 0x7f, 0x6b,
+ 0x0e, 0x7a, 0x95, 0xb9, 0x5c, 0x56, 0xfe, 0xc2,
+ 0x4d, 0xb7, 0xb4, 0xbd, 0x69, 0xb5, 0x41, 0x18,
+ 0x85, 0xa7, 0x47, 0x96, 0xe9, 0x25, 0x38, 0xfd,
+ 0xe7, 0x5f, 0xad, 0x44, 0xbb, 0x09, 0x53, 0x86,
+ 0x48, 0x5a, 0xf0, 0x57, 0x21, 0xef, 0xb1, 0x4f,
+ 0xa4, 0x48, 0xf6, 0xd9, 0x4d, 0x6d, 0xce, 0x24,
+ 0xaa, 0x32, 0x63, 0x60, 0x11, 0x3b, 0x30, 0xe6,
+ 0xa2, 0x5e, 0x7e, 0xd5, 0x83, 0xb1, 0xcf, 0x9a,
+ 0x27, 0xf9, 0x39, 0x43, 0x6a, 0x94, 0xf7, 0x67,
+ 0xc0, 0xa6, 0x94, 0x07, 0xd1, 0x9d, 0xa4, 0xe1,
+ 0xec, 0x17, 0x86, 0xeb, 0x6f, 0xa6, 0x49, 0x71,
+ 0x48, 0x5f, 0x70, 0x32, 0x22, 0xcb, 0x87, 0x55,
+ 0xe2, 0x6d, 0x13, 0x52, 0x33, 0xf0, 0xb7, 0xb3,
+ 0x40, 0xbe, 0xeb, 0x28, 0x2f, 0x18, 0xa2, 0x59,
+ 0x67, 0x47, 0xd2, 0x6b, 0x45, 0x8c, 0x55, 0x3e,
+ 0xa7, 0xe1, 0x46, 0x6c, 0x94, 0x11, 0xf1, 0xdf,
+ 0x82, 0x1f, 0x75, 0x0a, 0xad, 0x07, 0xd7, 0x53,
+ 0xca, 0x40, 0x05, 0x38, 0x8f, 0xcc, 0x50, 0x06,
+ 0x28, 0x2d, 0x16, 0x6a, 0xbc, 0x3c, 0xe7, 0xb5,
+ 0xe9, 0x8b, 0xa0, 0x6f, 0x44, 0x8c, 0x77, 0x3c,
+ 0x8e, 0xcc, 0x72, 0x04, 0x01, 0x00, 0x22, 0x02,
+ ];
+
+ let block = x86ni();
+ x86ni_init(&block, key[..]);
+
+ assert(block.rklen == 208);
+ assert(bytes::equal(expected_rounds[..], block.enc_rk[..RKLEN_192]));
+};
+
+
+// taken from fips-197.pdf Section A.3
+@test fn ni_enc_key_expand_256() void = {
+ const key: [32]u8 = [
+ 0x60, 0x3d, 0xeb, 0x10, 0x15, 0xca, 0x71, 0xbe,
+ 0x2b, 0x73, 0xae, 0xf0, 0x85, 0x7d, 0x77, 0x81,
+ 0x1f, 0x35, 0x2c, 0x07, 0x3b, 0x61, 0x08, 0xd7,
+ 0x2d, 0x98, 0x10, 0xa3, 0x09, 0x14, 0xdf, 0xf4,
+ ];
+
+ const expected_rounds: [_]u8 = [
+ 0x60, 0x3d, 0xeb, 0x10, 0x15, 0xca, 0x71, 0xbe,
+ 0x2b, 0x73, 0xae, 0xf0, 0x85, 0x7d, 0x77, 0x81,
+ 0x1f, 0x35, 0x2c, 0x07, 0x3b, 0x61, 0x08, 0xd7,
+ 0x2d, 0x98, 0x10, 0xa3, 0x09, 0x14, 0xdf, 0xf4,
+ 0x9b, 0xa3, 0x54, 0x11, 0x8e, 0x69, 0x25, 0xaf,
+ 0xa5, 0x1a, 0x8b, 0x5f, 0x20, 0x67, 0xfc, 0xde,
+ 0xa8, 0xb0, 0x9c, 0x1a, 0x93, 0xd1, 0x94, 0xcd,
+ 0xbe, 0x49, 0x84, 0x6e, 0xb7, 0x5d, 0x5b, 0x9a,
+ 0xd5, 0x9a, 0xec, 0xb8, 0x5b, 0xf3, 0xc9, 0x17,
+ 0xfe, 0xe9, 0x42, 0x48, 0xde, 0x8e, 0xbe, 0x96,
+ 0xb5, 0xa9, 0x32, 0x8a, 0x26, 0x78, 0xa6, 0x47,
+ 0x98, 0x31, 0x22, 0x29, 0x2f, 0x6c, 0x79, 0xb3,
+ 0x81, 0x2c, 0x81, 0xad, 0xda, 0xdf, 0x48, 0xba,
+ 0x24, 0x36, 0x0a, 0xf2, 0xfa, 0xb8, 0xb4, 0x64,
+ 0x98, 0xc5, 0xbf, 0xc9, 0xbe, 0xbd, 0x19, 0x8e,
+ 0x26, 0x8c, 0x3b, 0xa7, 0x09, 0xe0, 0x42, 0x14,
+ 0x68, 0x00, 0x7b, 0xac, 0xb2, 0xdf, 0x33, 0x16,
+ 0x96, 0xe9, 0x39, 0xe4, 0x6c, 0x51, 0x8d, 0x80,
+ 0xc8, 0x14, 0xe2, 0x04, 0x76, 0xa9, 0xfb, 0x8a,
+ 0x50, 0x25, 0xc0, 0x2d, 0x59, 0xc5, 0x82, 0x39,
+ 0xde, 0x13, 0x69, 0x67, 0x6c, 0xcc, 0x5a, 0x71,
+ 0xfa, 0x25, 0x63, 0x95, 0x96, 0x74, 0xee, 0x15,
+ 0x58, 0x86, 0xca, 0x5d, 0x2e, 0x2f, 0x31, 0xd7,
+ 0x7e, 0x0a, 0xf1, 0xfa, 0x27, 0xcf, 0x73, 0xc3,
+ 0x74, 0x9c, 0x47, 0xab, 0x18, 0x50, 0x1d, 0xda,
+ 0xe2, 0x75, 0x7e, 0x4f, 0x74, 0x01, 0x90, 0x5a,
+ 0xca, 0xfa, 0xaa, 0xe3, 0xe4, 0xd5, 0x9b, 0x34,
+ 0x9a, 0xdf, 0x6a, 0xce, 0xbd, 0x10, 0x19, 0x0d,
+ 0xfe, 0x48, 0x90, 0xd1, 0xe6, 0x18, 0x8d, 0x0b,
+ 0x04, 0x6d, 0xf3, 0x44, 0x70, 0x6c, 0x63, 0x1e,
+ ];
+
+ let block = x86ni();
+ x86ni_init(&block, key[..]);
+
+ assert(block.rklen == 240);
+ assert(bytes::equal(expected_rounds[..], block.enc_rk[..RKLEN_256]));
+};
+
+@test fn ni_test_encrypt_128() void = {
+ let key: [_]u8 = [
+ 0x2b, 0x7e, 0x15, 0x16, 0x28, 0xae, 0xd2, 0xa6,
+ 0xab, 0xf7, 0x15, 0x88, 0x09, 0xcf, 0x4f, 0x3c,
+ ];
+
+ let plain: [16]u8 = [
+ 0x32, 0x43, 0xf6, 0xa8, 0x88, 0x5a, 0x30, 0x8d,
+ 0x31, 0x31, 0x98, 0xa2, 0xe0, 0x37, 0x07, 0x34,
+ ];
+
+ const cipher: [16]u8 = [
+ 0x39, 0x25, 0x84, 0x1d, 0x02, 0xdc, 0x09, 0xfb,
+ 0xdc, 0x11, 0x85, 0x97, 0x19, 0x6a, 0x0b, 0x32,
+ ];
+
+ let result: [16]u8 = [0...];
+ let b = x86ni();
+
+ x86ni_init(&b, key[..]);
+ cipher::encrypt(&b, result[..], plain);
+
+ assert(bytes::equal(cipher, result));
+};
+
+@test fn ni_test_decrypt_128() void = {
+ const key: [_]u8 = [
+ 0x2b, 0x7e, 0x15, 0x16, 0x28, 0xae, 0xd2, 0xa6,
+ 0xab, 0xf7, 0x15, 0x88, 0x09, 0xcf, 0x4f, 0x3c,
+ ];
+
+ const plain: [16]u8 = [
+ 0x32, 0x43, 0xf6, 0xa8, 0x88, 0x5a, 0x30, 0x8d,
+ 0x31, 0x31, 0x98, 0xa2, 0xe0, 0x37, 0x07, 0x34,
+ ];
+
+ const cipher: [16]u8 = [
+ 0x39, 0x25, 0x84, 0x1d, 0x02, 0xdc, 0x09, 0xfb,
+ 0xdc, 0x11, 0x85, 0x97, 0x19, 0x6a, 0x0b, 0x32,
+ ];
+
+ let result: [16]u8 = [0...];
+ let b = x86ni();
+
+ x86ni_init(&b, key[..]);
+ cipher::decrypt(&b, result[..], cipher);
+ assert(bytes::equal(plain, result));
+};
+
+// fips-197.pdf Appendix C.1
+@test fn ni_test_example_vector1() void = {
+ const key: []u8 = [
+ 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
+ 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f,
+ ];
+
+ const plain: []u8 = [
+ 0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+ 0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+ ];
+
+ const cipher: []u8 = [
+ 0x69, 0xc4, 0xe0, 0xd8, 0x6a, 0x7b, 0x04, 0x30,
+ 0xd8, 0xcd, 0xb7, 0x80, 0x70, 0xb4, 0xc5, 0x5a,
+ ];
+
+ let result: [16]u8 = [0...];
+ let b = x86ni();
+
+ x86ni_init(&b, key[..]);
+
+ cipher::encrypt(&b, result[..], plain);
+ assert(bytes::equal(cipher, result));
+
+ cipher::decrypt(&b, result[..], cipher);
+ assert(bytes::equal(plain, result));
+};
+
+// fips-197.pdf Appendix C.2
+@test fn ni_test_example_vector2() void = {
+ const key: []u8 = [
+ 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
+ 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f,
+ 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17,
+ ];
+
+ const plain: []u8 = [
+ 0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+ 0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+ ];
+
+ const cipher: []u8 = [
+ 0xdd, 0xa9, 0x7c, 0xa4, 0x86, 0x4c, 0xdf, 0xe0,
+ 0x6e, 0xaf, 0x70, 0xa0, 0xec, 0x0d, 0x71, 0x91,
+ ];
+
+ let result: [16]u8 = [0...];
+ let b = x86ni();
+
+ x86ni_init(&b, key[..]);
+
+ cipher::encrypt(&b, result[..], plain);
+ assert(bytes::equal(cipher, result));
+
+ cipher::decrypt(&b, result[..], cipher);
+ assert(bytes::equal(plain, result));
+};
+
+// fips-197.pdf Appendix C.3
+@test fn ni_test_example_vector3() void = {
+ const key: []u8 = [
+ 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
+ 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f,
+ 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17,
+ 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f,
+ ];
+
+ const plain: []u8 = [
+ 0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+ 0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+ ];
+
+ const cipher: []u8 = [
+ 0x8e, 0xa2, 0xb7, 0xca, 0x51, 0x67, 0x45, 0xbf,
+ 0xea, 0xfc, 0x49, 0x90, 0x4b, 0x49, 0x60, 0x89,
+ ];
+
+ let result: [16]u8 = [0...];
+ let b = x86ni();
+
+ x86ni_init(&b, key[..]);
+
+ cipher::encrypt(&b, result[..], plain);
+ assert(bytes::equal(cipher, result));
+
+ cipher::decrypt(&b, result[..], cipher);
+ assert(bytes::equal(plain, result));
+};
diff --git a/crypto/aes/+x86_64/ni.ha b/crypto/aes/+x86_64/ni.ha
@@ -0,0 +1,64 @@
+use bytes;
+use crypto::cipher;
+use rt;
+
+def RKLEN_128: size = 176;
+def RKLEN_192: size = 208;
+def RKLEN_256: size = 240;
+
+export type x86ni_block = struct {
+ cipher::block,
+ enc_rk: [RKLEN_256]u8,
+ dec_rk: [RKLEN_256]u8,
+ rklen: u8,
+};
+
+// Checks if the native AES interface is available.
+export fn x86ni_available() bool = {
+ return rt::cpuid_hasflags(0, rt::cpuid_ecxflags::AES);
+};
+
+// Returns a native AES [[crypto::cipher::block]] implementation for x86_64
+// CPUs supporting AES-NI.
+//
+// The caller must call [[x86ni_init]] to add a key to the cipher before using
+// the cipher, and must call [[crypto::cipher::finish]] when they are finished
+// using the cipher to securely erase any secret data stored in the cipher
+// state.
+export fn x86ni() x86ni_block = {
+ return x86ni_block {
+ encrypt = &x86ni_encrypt,
+ decrypt = &x86ni_decrypt,
+ finish = &x86ni_finish,
+ ...
+ };
+};
+
+export fn x86ni_init(b: *x86ni_block, key: []u8) void = {
+ assert(len(key) == 16 || len(key) == 24 || len(key) == 32,
+ "Invalid aes key length");
+
+ b.rklen = x86ni_keyexp(key[..], b.enc_rk[..], b.dec_rk[..]);
+};
+
+fn x86ni_encrypt(block: *cipher::block, dest: []u8, src: []u8) void = {
+ let b = block: *x86ni_block;
+ x86ni_asencrypt(b.enc_rk[..b.rklen], dest, src);
+};
+
+fn x86ni_decrypt(block: *cipher::block, dest: []u8, src: []u8) void = {
+ let b = block: *x86ni_block;
+ x86ni_asdecrypt(b.dec_rk[..b.rklen], dest, src);
+};
+
+fn x86ni_finish(block: *cipher::block) void = {
+ let b = block: *x86ni_block;
+ bytes::zero(b.enc_rk[..]);
+ bytes::zero(b.dec_rk[..]);
+};
+
+// Expands encryption and decryption key and returns the size of the round keys.
+export fn x86ni_keyexp(key: []u8, enc_rk: []u8, dec_rk: []u8) u8;
+export fn x86ni_asencrypt(key_exp: []u8, dest: []u8, src: []u8) void;
+export fn x86ni_asdecrypt(key_exp: []u8, dest: []u8, src: []u8) void;
+
diff --git a/crypto/aes/+x86_64/ni_native.s b/crypto/aes/+x86_64/ni_native.s
@@ -0,0 +1,500 @@
+.global crypto.aes.x86ni_keyexp
+.type crypto.aes.x86ni_keyexp,@function
+crypto.aes.x86ni_keyexp:
+ pushq %rbp
+ mov %rsp, %rbp
+
+ pushq %rbx
+ pushq %rcx
+ pushq %rdx
+
+ movq 0x10(%rbp), %rbx # &key
+ movq 0x18(%rbp), %rax # keylen
+
+ movq 0x28(%rbp), %rcx # &enk_rk
+
+ mov $0x18, %rdx
+ cmp %rax, %rdx
+ je enc_key_192
+ jle enc_key_256
+
+enc_key_128:
+ movdqu (%rbx), %xmm1
+ movdqu %xmm1, (%rcx)
+ aeskeygenassist $0x1, %xmm1, %xmm2
+ call key_expand_128
+ movdqu %xmm1, 0x10(%rcx)
+ aeskeygenassist $0x2, %xmm1, %xmm2
+ call key_expand_128
+ movdqu %xmm1, 0x20(%rcx)
+ aeskeygenassist $0x4, %xmm1, %xmm2
+ call key_expand_128
+ movdqu %xmm1, 0x30(%rcx)
+ aeskeygenassist $0x8, %xmm1, %xmm2
+ call key_expand_128
+ movdqu %xmm1, 0x40(%rcx)
+ aeskeygenassist $0x10, %xmm1, %xmm2
+ call key_expand_128
+ movdqu %xmm1, 0x50(%rcx)
+ aeskeygenassist $0x20, %xmm1, %xmm2
+ call key_expand_128
+ movdqu %xmm1, 0x60(%rcx)
+ aeskeygenassist $0x40, %xmm1, %xmm2
+ call key_expand_128
+ movdqu %xmm1, 0x70(%rcx)
+ aeskeygenassist $0x80, %xmm1, %xmm2
+ call key_expand_128
+ movdqu %xmm1, 0x80(%rcx)
+ aeskeygenassist $0x1b, %xmm1, %xmm2
+ call key_expand_128
+ movdqu %xmm1, 0x90(%rcx)
+ aeskeygenassist $0x36, %xmm1, %xmm2
+ call key_expand_128
+ movdqu %xmm1, 0xa0(%rcx)
+
+ jmp dec_key
+key_expand_128:
+ vpslldq $0x4, %xmm1, %xmm3
+ pxor %xmm3, %xmm1
+ vpslldq $0x4, %xmm1, %xmm3
+ pxor %xmm3, %xmm1
+ vpslldq $0x4, %xmm1, %xmm3
+ pxor %xmm3, %xmm1
+
+ pshufd $0xff, %xmm2, %xmm2
+ pxor %xmm2, %xmm1
+ ret
+
+enc_key_192:
+ movdqu (%rbx), %xmm1
+ movdqu 0x10(%rbx), %xmm3
+
+ movdqu %xmm1, (%rcx)
+ movdqu %xmm3, %xmm5
+
+ aeskeygenassist $0x1, %xmm3, %xmm2
+ call key_expand_192
+ shufpd $0, %xmm1, %xmm5
+ movdqu %xmm5, 0x10(%rcx)
+ movdqu %xmm1, %xmm6
+ shufpd $1, %xmm3, %xmm6
+ movdqu %xmm6, 0x20(%rcx)
+
+ aeskeygenassist $0x2, %xmm3, %xmm2
+ call key_expand_192
+ movdqu %xmm1, 0x30(%rcx)
+ movdqu %xmm3, %xmm5
+
+ aeskeygenassist $0x4, %xmm3, %xmm2
+ call key_expand_192
+ shufpd $0, %xmm1, %xmm5
+ movdqu %xmm5, 0x40(%rcx)
+ movdqu %xmm1, %xmm6
+ shufpd $1, %xmm3, %xmm6
+ movdqu %xmm6, 0x50(%rcx)
+
+ aeskeygenassist $0x8, %xmm3, %xmm2
+ call key_expand_192
+ movdqu %xmm1, 0x60(%rcx)
+ movdqu %xmm3, %xmm5
+
+ aeskeygenassist $0x10, %xmm3, %xmm2
+ call key_expand_192
+ shufpd $0, %xmm1, %xmm5
+ movdqu %xmm5, 0x70(%rcx)
+ movdqu %xmm1, %xmm6
+ shufpd $1, %xmm3, %xmm6
+ movdqu %xmm6, 0x80(%rcx)
+
+ aeskeygenassist $0x20, %xmm3, %xmm2
+ call key_expand_192
+ movdqu %xmm1, 0x90(%rcx)
+ movdqu %xmm3, %xmm5
+
+ aeskeygenassist $0x40, %xmm3, %xmm2
+ call key_expand_192
+ shufpd $0, %xmm1, %xmm5
+ movdqu %xmm5, 0xa0(%rcx)
+ movdqu %xmm1, %xmm6
+ shufpd $1, %xmm3, %xmm6
+ movdqu %xmm6, 0xb0(%rcx)
+
+ aeskeygenassist $0x80, %xmm3, %xmm2
+ call key_expand_192
+ movdqu %xmm1, 0xc0(%rcx)
+ movdqu %xmm3, %xmm5
+
+ jmp dec_key
+
+key_expand_192:
+ vpslldq $0x4, %xmm1, %xmm4
+ pxor %xmm4, %xmm1
+ vpslldq $0x4, %xmm1, %xmm4
+ pxor %xmm4, %xmm1
+ vpslldq $0x4, %xmm1, %xmm4
+ pxor %xmm4, %xmm1
+
+ pshufd $0x55, %xmm2, %xmm2
+ pxor %xmm2, %xmm1
+
+ pshufd $0xff, %xmm1, %xmm2
+ vpslldq $0x4, %xmm3, %xmm4
+
+ pxor %xmm4, %xmm3
+ pxor %xmm2, %xmm3
+
+ ret
+
+enc_key_256:
+ movdqu (%rbx), %xmm1
+ movdqu 0x10(%rbx), %xmm3
+
+ movdqu %xmm1, (%rcx)
+ movdqu %xmm3, 0x10(%rcx)
+
+ aeskeygenassist $0x1, %xmm3, %xmm2
+ call key_expand_256_a
+ movdqu %xmm1, 0x20(%rcx)
+ aeskeygenassist $0x0, %xmm1, %xmm2
+ call key_expand_256_b
+ movdqu %xmm3, 0x30(%rcx)
+ aeskeygenassist $0x2, %xmm3, %xmm2
+ call key_expand_256_a
+ movdqu %xmm1, 0x40(%rcx)
+ aeskeygenassist $0x0, %xmm1, %xmm2
+ call key_expand_256_b
+ movdqu %xmm3, 0x50(%rcx)
+ aeskeygenassist $0x4, %xmm3, %xmm2
+ call key_expand_256_a
+ movdqu %xmm1, 0x60(%rcx)
+ aeskeygenassist $0x0, %xmm1, %xmm2
+ call key_expand_256_b
+ movdqu %xmm3, 0x70(%rcx)
+ aeskeygenassist $0x8, %xmm3, %xmm2
+ call key_expand_256_a
+ movdqu %xmm1, 0x80(%rcx)
+ aeskeygenassist $0x0, %xmm1, %xmm2
+ call key_expand_256_b
+ movdqu %xmm3, 0x90(%rcx)
+ aeskeygenassist $0x10, %xmm3, %xmm2
+ call key_expand_256_a
+ movdqu %xmm1, 0xa0(%rcx)
+ aeskeygenassist $0x0, %xmm1, %xmm2
+ call key_expand_256_b
+ movdqu %xmm3, 0xb0(%rcx)
+ aeskeygenassist $0x20, %xmm3, %xmm2
+ call key_expand_256_a
+ movdqu %xmm1, 0xc0(%rcx)
+ aeskeygenassist $0x0, %xmm1, %xmm2
+ call key_expand_256_b
+ movdqu %xmm3, 0xd0(%rcx)
+ aeskeygenassist $0x40, %xmm3, %xmm2
+ call key_expand_256_a
+ movdqu %xmm1, 0xe0(%rcx)
+ jmp dec_key
+
+key_expand_256_a:
+ movdqa %xmm1, %xmm4
+ pslldq $4, %xmm4
+ pxor %xmm4, %xmm1
+ pslldq $4, %xmm4
+ pxor %xmm4, %xmm1
+ pslldq $4, %xmm4
+ pxor %xmm4, %xmm1
+
+ pshufd $0xff, %xmm2, %xmm2
+ pxor %xmm2, %xmm1
+
+ ret
+
+key_expand_256_b:
+ movdqa %xmm3, %xmm4
+ pslldq $4, %xmm4
+ pxor %xmm4, %xmm3
+ pslldq $4, %xmm4
+ pxor %xmm4, %xmm3
+ pslldq $4, %xmm4
+ pxor %xmm4, %xmm3
+
+ pshufd $0xaa, %xmm2, %xmm2
+ pxor %xmm2, %xmm3
+
+ ret
+
+dec_key:
+ movq 0x40(%rbp), %rdx # &dec_rk
+
+ # store key in reverse order, therefore add rklen to rk pointer
+ mov $0x18, %rbx
+ je rklen_dec_key_192
+ jle rklen_dec_key_256
+ add $160, %rdx
+ jmp dec_key_start
+rklen_dec_key_192:
+ add $192, %rdx
+ jmp dec_key_start
+rklen_dec_key_256:
+ add $224, %rdx
+
+dec_key_start:
+ movdqu 0x0(%rcx), %xmm1
+ movdqu %xmm1, 0x0(%rdx)
+
+ movdqu 0x10(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0x10(%rdx)
+ movdqu 0x20(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0x20(%rdx)
+ movdqu 0x30(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0x30(%rdx)
+ movdqu 0x40(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0x40(%rdx)
+ movdqu 0x50(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0x50(%rdx)
+ movdqu 0x60(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0x60(%rdx)
+ movdqu 0x70(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0x70(%rdx)
+ movdqu 0x80(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0x80(%rdx)
+ movdqu 0x90(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0x90(%rdx)
+
+ cmp %rax, %rbx
+ je dec_key_192
+ jle dec_key_256
+
+ movdqu 0xa0(%rcx), %xmm1
+ movdqu %xmm1, -0xa0(%rdx)
+
+ # return rklen
+ movl $176, %eax
+
+ jmp key_exp_end
+
+dec_key_192:
+ movdqu 0xa0(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0xa0(%rdx)
+ movdqu 0xb0(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0xb0(%rdx)
+
+
+ movdqu 0xc0(%rcx), %xmm1
+ movdqu %xmm1, -0xc0(%rdx)
+
+ # return rklen
+ movl $208, %eax
+ jmp key_exp_end
+dec_key_256:
+ movdqu 0xa0(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0xa0(%rdx)
+ movdqu 0xb0(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0xb0(%rdx)
+ movdqu 0xc0(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0xc0(%rdx)
+ movdqu 0xd0(%rcx), %xmm1
+ aesimc %xmm1, %xmm1
+ movdqu %xmm1, -0xd0(%rdx)
+
+
+ movdqu 0xe0(%rcx), %xmm1
+ movdqu %xmm1, -0xe0(%rdx)
+
+ # return rklen
+ movl $240, %eax
+
+key_exp_end:
+ pxor %xmm0, %xmm0
+ pxor %xmm1, %xmm1
+ pxor %xmm2, %xmm2
+ pxor %xmm3, %xmm3
+ pxor %xmm4, %xmm4
+ pxor %xmm5, %xmm5
+ pxor %xmm6, %xmm6
+
+ popq %rdx
+ popq %rcx
+ popq %rbx
+
+ leave
+ ret
+
+.global crypto.aes.x86ni_asencrypt
+.type crypto.aes.x86ni_asencrypt,@function
+crypto.aes.x86ni_asencrypt:
+ pushq %rbp
+ mov %rsp, %rbp
+ pushq %rbx
+ pushq %rcx
+ pushq %rdx
+
+ movq 0x10(%rbp), %rbx # &rk
+ movq 0x18(%rbp), %rax # rklen
+
+ movq 0x28(%rbp), %rcx # &dest
+ movq 0x40(%rbp), %rdx # &src
+
+ movdqu (%rdx), %xmm0
+ movdqu (%rbx), %xmm1
+ pxor %xmm1, %xmm0
+
+ movdqu 0x10(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0x20(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0x30(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0x40(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0x50(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0x60(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0x70(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0x80(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0x90(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+
+ mov $208, %rdx
+ cmp %rax, %rdx
+ jl encrypt_256
+ je encrypt_192
+
+ movdqu 0xa0(%rbx), %xmm1
+ aesenclast %xmm1, %xmm0
+ jmp encrypt_end
+
+encrypt_192:
+ movdqu 0xa0(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0xb0(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0xc0(%rbx), %xmm1
+ aesenclast %xmm1, %xmm0
+ jmp encrypt_end
+
+encrypt_256:
+ movdqu 0xa0(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0xb0(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0xc0(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0xd0(%rbx), %xmm1
+ aesenc %xmm1, %xmm0
+ movdqu 0xe0(%rbx), %xmm1
+ aesenclast %xmm1, %xmm0
+ jmp encrypt_end
+
+encrypt_end:
+
+ movdqu %xmm0, (%rcx)
+
+ pxor %xmm0, %xmm0
+ pxor %xmm1, %xmm1
+
+ popq %rdx
+ popq %rcx
+ popq %rbx
+
+ leave
+ ret
+
+.global crypto.aes.x86ni_asdecrypt
+.type crypto.aes.x86ni_asdecrypt,@function
+crypto.aes.x86ni_asdecrypt:
+ pushq %rbp
+ mov %rsp, %rbp
+
+ pushq %rbx
+ pushq %rcx
+ pushq %rdx
+
+ movq 0x10(%rbp), %rbx # &rk
+ movq 0x18(%rbp), %rax # rklen
+
+ movq 0x28(%rbp), %rcx # &dest
+ movq 0x40(%rbp), %rdx # &src
+
+ movdqu (%rdx), %xmm0
+ movdqu (%rbx), %xmm1
+ pxor %xmm1, %xmm0
+
+ movdqu 0x10(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0x20(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0x30(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0x40(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0x50(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0x60(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0x70(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0x80(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0x90(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+
+ mov $208, %rdx
+ cmp %rax, %rdx
+ je decrypt_192
+ jl decrypt_256
+
+ movdqu 0xa0(%rbx), %xmm1
+ aesdeclast %xmm1, %xmm0
+ jmp decrypt_end
+
+decrypt_192:
+ movdqu 0xa0(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0xb0(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0xc0(%rbx), %xmm1
+ aesdeclast %xmm1, %xmm0
+ jmp decrypt_end
+
+decrypt_256:
+ movdqu 0xa0(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0xb0(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0xc0(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0xd0(%rbx), %xmm1
+ aesdec %xmm1, %xmm0
+ movdqu 0xe0(%rbx), %xmm1
+ aesdeclast %xmm1, %xmm0
+ jmp decrypt_end
+
+decrypt_end:
+ movdqu %xmm0, (%rcx)
+
+ pxor %xmm0, %xmm0
+ pxor %xmm1, %xmm1
+
+ popq %rdx
+ popq %rcx
+ popq %rbx
+
+ leave
+ ret
+
diff --git a/scripts/gen-stdlib b/scripts/gen-stdlib
@@ -231,10 +231,13 @@ crypto_aes() {
if [ $testing -eq 0 ]
then
gensrcs_crypto_aes
- gen_ssa crypto::aes crypto::cipher crypto::math endian
+ gen_ssa crypto::aes bytes crypto::cipher crypto::math endian rt
else
- gensrcs_crypto_aes ct64+test.ha cbc+test.ha ctr+test.ha
- gen_ssa crypto::aes bytes crypto::cipher crypto::math endian
+ gensrcs_crypto_aes \
+ ct64+test.ha \
+ cbc+test.ha \
+ ctr+test.ha
+ gen_ssa crypto::aes bytes crypto::cipher crypto::math endian rt
fi
}
diff --git a/stdlib.mk b/stdlib.mk
@@ -742,7 +742,7 @@ $(HARECACHE)/crypto/crypto-any.ssa: $(stdlib_crypto_any_srcs) $(stdlib_rt) $(std
stdlib_crypto_aes_any_srcs= \
$(STDLIB)/crypto/aes/aes_ct64.ha
-$(HARECACHE)/crypto/aes/crypto_aes-any.ssa: $(stdlib_crypto_aes_any_srcs) $(stdlib_rt) $(stdlib_crypto_cipher_$(PLATFORM)) $(stdlib_crypto_math_$(PLATFORM)) $(stdlib_endian_$(PLATFORM))
+$(HARECACHE)/crypto/aes/crypto_aes-any.ssa: $(stdlib_crypto_aes_any_srcs) $(stdlib_rt) $(stdlib_bytes_$(PLATFORM)) $(stdlib_crypto_cipher_$(PLATFORM)) $(stdlib_crypto_math_$(PLATFORM)) $(stdlib_endian_$(PLATFORM)) $(stdlib_rt_$(PLATFORM))
@printf 'HAREC \t$@\n'
@mkdir -p $(HARECACHE)/crypto/aes
@HARECACHE=$(HARECACHE) $(HAREC) $(HAREFLAGS) -o $@ -Ncrypto::aes \
@@ -2654,7 +2654,7 @@ testlib_crypto_aes_any_srcs= \
$(STDLIB)/crypto/aes/cbc+test.ha \
$(STDLIB)/crypto/aes/ctr+test.ha
-$(TESTCACHE)/crypto/aes/crypto_aes-any.ssa: $(testlib_crypto_aes_any_srcs) $(testlib_rt) $(testlib_bytes_$(PLATFORM)) $(testlib_crypto_cipher_$(PLATFORM)) $(testlib_crypto_math_$(PLATFORM)) $(testlib_endian_$(PLATFORM))
+$(TESTCACHE)/crypto/aes/crypto_aes-any.ssa: $(testlib_crypto_aes_any_srcs) $(testlib_rt) $(testlib_bytes_$(PLATFORM)) $(testlib_crypto_cipher_$(PLATFORM)) $(testlib_crypto_math_$(PLATFORM)) $(testlib_endian_$(PLATFORM)) $(testlib_rt_$(PLATFORM))
@printf 'HAREC \t$@\n'
@mkdir -p $(TESTCACHE)/crypto/aes
@HARECACHE=$(TESTCACHE) $(HAREC) $(TESTHAREFLAGS) -o $@ -Ncrypto::aes \
