dark/zerotiertspu
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Cleanup and warning removal.

Browse source
This commit is contained in:
Adam Ierymenko 2019-09-13 16:24:02 -07:00
parent 063a16fcd6
commit 9f9032ae36
No known key found for this signature in database
GPG key ID: C8877CF2D7A5D7F3
10 changed files with 55 additions and 108 deletions
Download patch file Download diff file Expand all files Collapse all files

83
node/AES-aesni.c
View file

@ -19,20 +19,14 @@
#if (defined(__amd64) || defined(__amd64__) || defined(__x86_64) || defined(__x86_64__) || defined(__AMD64) || defined(__AMD64__) || defined(_M_X64))
#include <stdint.h>
#include <string.h>
#include <stdlib.h>
#include <stdio.h>
#include <wmmintrin.h>
#include <emmintrin.h>
#include <smmintrin.h>
#include <immintrin.h>
/* #define register */
#define ZT_AES_CTR_AESNI_ROUND(kk) c0 = _mm_aesenc_si128(c0,kk); c1 = _mm_aesenc_si128(c1,kk); c2 = _mm_aesenc_si128(c2,kk); c3 = _mm_aesenc_si128(c3,kk);
void zt_crypt_ctr_aesni(const __m128i key[14],const uint8_t iv[16],const uint8_t *in,unsigned int len,uint8_t *out)
{
_mm_prefetch(in,_MM_HINT_NTA);
/* Because our CTR supports full 128-bit nonces, we must do a full 128-bit (big-endian)
* increment to be compatible with canonical NIST-certified CTR implementations. That's
* because it's possible to have a lot of bit saturation in the least significant 64
@ -55,82 +49,61 @@ void zt_crypt_ctr_aesni(const __m128i key[14],const uint8_t iv[16],const uint8_t
register __m128i ctr3 = _mm_shuffle_epi8(_mm_add_epi64(ctr0,_mm_set_epi64x((long long)(notctr0msq < 3ULL),3LL)),swap128);
ctr0 = _mm_shuffle_epi8(ctr0,swap128);
register __m128i k0 = key[0];
register __m128i k1 = key[1];
while (len >= 64) {
_mm_prefetch(in + 64,_MM_HINT_NTA);
register __m128i ka = key[0];
register __m128i c0 = _mm_xor_si128(ctr0,ka);
register __m128i ka = key[2];
register __m128i c0 = _mm_xor_si128(ctr0,k0);
register __m128i c1 = _mm_xor_si128(ctr1,k0);
register __m128i c2 = _mm_xor_si128(ctr2,k0);
register __m128i c3 = _mm_xor_si128(ctr3,k0);
ctr0 = _mm_shuffle_epi8(ctr0,swap128);
notctr0msq = ~((uint64_t)_mm_extract_epi64(ctr0,0));
register __m128i c1 = _mm_xor_si128(ctr1,ka);
register __m128i c2 = _mm_xor_si128(ctr2,ka);
register __m128i c3 = _mm_xor_si128(ctr3,ka);
register __m128i kb = key[1];
ctr1 = _mm_shuffle_epi8(_mm_add_epi64(ctr0,_mm_set_epi64x((long long)(notctr0msq < 5ULL),5LL)),swap128);
ctr2 = _mm_shuffle_epi8(_mm_add_epi64(ctr0,_mm_set_epi64x((long long)(notctr0msq < 6ULL),6LL)),swap128);
register __m128i kc = key[2];
ctr3 = _mm_shuffle_epi8(_mm_add_epi64(ctr0,_mm_set_epi64x((long long)(notctr0msq < 7ULL),7LL)),swap128);
ctr0 = _mm_shuffle_epi8(_mm_add_epi64(ctr0,_mm_set_epi64x((long long)(notctr0msq < 4ULL),4LL)),swap128);
register __m128i kd = key[3];
#define ZT_AES_CTR_AESNI_ROUND(kk) \
c0 = _mm_aesenc_si128(c0,kk); \
c1 = _mm_aesenc_si128(c1,kk); \
c2 = _mm_aesenc_si128(c2,kk); \
c3 = _mm_aesenc_si128(c3,kk);
ka = key[4];
ZT_AES_CTR_AESNI_ROUND(kb);
kb = key[5];
ZT_AES_CTR_AESNI_ROUND(kc);
kc = key[6];
ZT_AES_CTR_AESNI_ROUND(kd);
kd = key[7];
register __m128i kb = key[3];
ZT_AES_CTR_AESNI_ROUND(k1);
register __m128i kc = key[4];
ZT_AES_CTR_AESNI_ROUND(ka);
ka = key[8];
register __m128i kd = key[5];
ZT_AES_CTR_AESNI_ROUND(kb);
kb = key[9];
ka = key[6];
ZT_AES_CTR_AESNI_ROUND(kc);
kc = key[10];
kb = key[7];
ZT_AES_CTR_AESNI_ROUND(kd);
kd = key[11];
kc = key[8];
ZT_AES_CTR_AESNI_ROUND(ka);
ka = key[12];
kd = key[9];
ZT_AES_CTR_AESNI_ROUND(kb);
kb = key[13];
ka = key[10];
ZT_AES_CTR_AESNI_ROUND(kc);
kc = key[14];
kb = key[11];
ZT_AES_CTR_AESNI_ROUND(kd);
kc = key[12];
ZT_AES_CTR_AESNI_ROUND(ka);
kd = key[13];
ZT_AES_CTR_AESNI_ROUND(kb);
#undef ZT_AES_CTR_AESNI_ROUND
register __m128i d0 = _mm_loadu_si128((const __m128i *)in);
register __m128i d1 = _mm_loadu_si128((const __m128i *)(in + 16));
register __m128i d2 = _mm_loadu_si128((const __m128i *)(in + 32));
register __m128i d3 = _mm_loadu_si128((const __m128i *)(in + 48));
c0 = _mm_aesenclast_si128(c0,kc);
c1 = _mm_aesenclast_si128(c1,kc);
c2 = _mm_aesenclast_si128(c2,kc);
c3 = _mm_aesenclast_si128(c3,kc);
d0 = _mm_xor_si128(d0,c0);
d1 = _mm_xor_si128(d1,c1);
d2 = _mm_xor_si128(d2,c2);
d3 = _mm_xor_si128(d3,c3);
_mm_storeu_si128((__m128i *)out,d0);
_mm_storeu_si128((__m128i *)(out + 16),d1);
_mm_storeu_si128((__m128i *)(out + 32),d2);
_mm_storeu_si128((__m128i *)(out + 48),d3);
ka = key[14];
ZT_AES_CTR_AESNI_ROUND(kc);
ZT_AES_CTR_AESNI_ROUND(kd);
_mm_storeu_si128((__m128i *)out,_mm_xor_si128(_mm_loadu_si128((const __m128i *)in),_mm_aesenclast_si128(c0,ka)));
_mm_storeu_si128((__m128i *)(out + 16),_mm_xor_si128(_mm_loadu_si128((const __m128i *)(in + 16)),_mm_aesenclast_si128(c1,ka)));
_mm_storeu_si128((__m128i *)(out + 32),_mm_xor_si128(_mm_loadu_si128((const __m128i *)(in + 32)),_mm_aesenclast_si128(c2,ka)));
_mm_storeu_si128((__m128i *)(out + 48),_mm_xor_si128(_mm_loadu_si128((const __m128i *)(in + 48)),_mm_aesenclast_si128(c3,ka)));
in += 64;
out += 64;
len -= 64;
}
register __m128i k0 = key[0];
register __m128i k1 = key[1];
register __m128i k2 = key[2];
register __m128i k3 = key[3];
register __m128i k4 = key[4];
register __m128i k5 = key[5];
register __m128i k6 = key[6];
register __m128i k7 = key[7];
/* not enough XMM registers for all of them, but it helps slightly... */
while (len >= 16) {
register __m128i c0 = _mm_xor_si128(ctr0,k0);