code: drawterm

Download patch

ref: b4d00e3be37d68ea5976a23ee15d1d35c8ffa89a
parent: a50be99567d1b6018a0a1f019a51c0bc7a0f11ff
author: cinap_lenrek <cinap_lenrek@felloff.net>
date: Sun Jan 15 14:02:22 EST 2017 

libsec: sync with 9front


--- a/include/libsec.h
+++ b/include/libsec.h
@@ -89,6 +89,7 @@
 	ChachaBsize=	64,
 	ChachaKeylen=	256/8,
 	ChachaIVlen=	96/8,
+	XChachaIVlen=	192/8,
 };
 
 typedef struct Chachastate Chachastate;
@@ -103,6 +104,7 @@
 			u32int	iv[3];
 		};
 	};
+	u32int	xkey[8];
 	int	rounds;
 	int	ivwords;
 };
@@ -113,6 +115,8 @@
 void	chacha_encrypt(uchar*, ulong, Chachastate*);
 void	chacha_encrypt2(uchar*, uchar*, ulong, Chachastate*);
 
+void	hchacha(uchar h[32], uchar *key, ulong keylen, uchar nonce[16], int rounds);
+
 void	ccpoly_encrypt(uchar *dat, ulong ndat, uchar *aad, ulong naad, uchar tag[16], Chachastate *cs);
 int	ccpoly_decrypt(uchar *dat, ulong ndat, uchar *aad, ulong naad, uchar tag[16], Chachastate *cs);
 
@@ -131,7 +135,7 @@
 struct Salsastate
 {
 	u32int	input[16];
-	u32int	key[8];
+	u32int	xkey[8];
 	int	rounds;
 	int	ivwords;
 };
@@ -142,6 +146,8 @@
 void	salsa_encrypt(uchar*, ulong, Salsastate*);
 void	salsa_encrypt2(uchar*, uchar*, ulong, Salsastate*);
 
+void	salsa_core(u32int in[16], u32int out[16], int rounds);
+
 void	hsalsa(uchar h[32], uchar *key, ulong keylen, uchar nonce[16], int rounds);
 
 /*
@@ -485,7 +491,7 @@
 	int inf;
 	mpint *x;
 	mpint *y;
-	mpint *z;
+	mpint *z;	/* nil when using affine coordinates */
 } ECpoint;
 
 typedef ECpoint ECpub;
@@ -563,6 +569,11 @@
 /* password-based key derivation function 2 (rfc2898) */
 void pbkdf2_x(uchar *p, ulong plen, uchar *s, ulong slen, ulong rounds, uchar *d, ulong dlen,
 	DigestState* (*x)(uchar*, ulong, uchar*, ulong, uchar*, DigestState*), int xlen);
+
+/* scrypt password-based key derivation function */
+char* scrypt(uchar *p, ulong plen, uchar *s, ulong slen,
+	ulong N, ulong R, ulong P,
+	uchar *d, ulong dlen);
 
 /* hmac-based key derivation function (rfc5869) */
 void hkdf_x(uchar *salt, ulong nsalt, uchar *info, ulong ninfo, uchar *key, ulong nkey, uchar *d, ulong dlen,
--- a/libsec/Makefile
+++ b/libsec/Makefile
@@ -58,6 +58,7 @@
 	rsagen.$O\
 	rsaprivtopub.$O\
 	salsa.$O\
+	scrypt.$O\
 	secp256k1.$O\
 	secp256r1.$O\
 	secp384r1.$O\
--- a/libsec/chacha.c
+++ b/libsec/chacha.c
@@ -15,13 +15,13 @@
 };
 
 /* little-endian data order */
-#define GET4(p)	((((((p)[3]<<8) | (p)[2])<<8) | (p)[1])<<8 | (p)[0])
-#define PUT4(p, v)	(((p)[0]=v), (v>>=8), ((p)[1]=v), (v>>=8), ((p)[2]=v), (v>>=8), ((p)[3]=v))
+#define	GET4(p)		((p)[0]|((p)[1]<<8)|((p)[2]<<16)|((p)[3]<<24))
+#define	PUT4(p,v)	(p)[0]=(v);(p)[1]=(v)>>8;(p)[2]=(v)>>16;(p)[3]=(v)>>24
 
 #define ROTATE(v,c) ((u32int)((v) << (c)) | ((v) >> (32 - (c))))
 
 #define QUARTERROUND(ia,ib,ic,id) { \
-	u32int a, b, c, d, t;\
+	u32int a, b, c, d, t; \
 	a = x[ia]; b = x[ib]; c = x[ic]; d = x[id]; \
 	a += b; t = d^a; d = ROTATE(t,16); \
 	c += d; t = b^c; b = ROTATE(t,12); \
@@ -32,12 +32,9 @@
 
 #define ENCRYPT(s, x, y, d) {\
 	u32int v; \
-	uchar *sp, *dp; \
-	sp = (s); \
-	v = GET4(sp); \
+	v = GET4(s); \
 	v ^= (x)+(y); \
-	dp = (d); \
-	PUT4(dp, v); \
+	PUT4(d, v); \
 }
 
 static uchar sigma[16] = "expand 32-byte k";
@@ -57,7 +54,8 @@
 {
 	if(keylen != 256/8 && keylen != 128/8)
 		sysfatal("invalid chacha key length");
-	if(ivlen != 96/8 && ivlen != 64/8)
+	if(ivlen != 64/8 && ivlen != 96/8
+	&& ivlen != 128/8 && ivlen != 192/8)	/* hchacha, xchacha */
 		sysfatal("invalid chacha iv length");
 	if(rounds == 0)
 		rounds = 20;
@@ -70,7 +68,16 @@
 		load(&s->input[4], key, 4);
 		load(&s->input[8], key, 4);
 	}
-	s->ivwords = ivlen/sizeof(u32int);
+	s->xkey[0] = s->input[4];
+	s->xkey[1] = s->input[5];
+	s->xkey[2] = s->input[6];
+	s->xkey[3] = s->input[7];
+	s->xkey[4] = s->input[8];
+	s->xkey[5] = s->input[9];
+	s->xkey[6] = s->input[10];
+	s->xkey[7] = s->input[11];
+
+	s->ivwords = ivlen/4;
 	s->input[12] = 0;
 	s->input[13] = 0;
 	if(iv == nil){
@@ -80,9 +87,88 @@
 		chacha_setiv(s, iv);
 }
 
+static void
+dorounds(u32int x[Blockwords], int rounds)
+{
+	for(; rounds > 0; rounds -= 2) {
+		QUARTERROUND(0, 4, 8,12)
+		QUARTERROUND(1, 5, 9,13)
+		QUARTERROUND(2, 6,10,14)
+		QUARTERROUND(3, 7,11,15)
+
+		QUARTERROUND(0, 5,10,15)
+		QUARTERROUND(1, 6,11,12)
+		QUARTERROUND(2, 7, 8,13)
+		QUARTERROUND(3, 4, 9,14)
+	}
+}
+
+static void
+hchachablock(uchar h[32], Chachastate *s)
+{
+	u32int x[16];
+
+	x[0] = s->input[0];
+	x[1] = s->input[1];
+	x[2] = s->input[2];
+	x[3] = s->input[3];
+	x[4] = s->input[4];
+	x[5] = s->input[5];
+	x[6] = s->input[6];
+	x[7] = s->input[7];
+	x[8] = s->input[8];
+	x[9] = s->input[9];
+	x[10] = s->input[10];
+	x[11] = s->input[11];
+	x[12] = s->input[12];
+	x[13] = s->input[13];
+	x[14] = s->input[14];
+	x[15] = s->input[15];
+
+	dorounds(x, s->rounds);
+
+	PUT4(h+0*4, x[0]);
+	PUT4(h+1*4, x[1]);
+	PUT4(h+2*4, x[2]);
+	PUT4(h+3*4, x[3]);
+	PUT4(h+4*4, x[12]);
+	PUT4(h+5*4, x[13]);
+	PUT4(h+6*4, x[14]);
+	PUT4(h+7*4, x[15]);
+}
+
 void
 chacha_setiv(Chachastate *s, uchar *iv)
 {
+	if(s->ivwords == 192/32){
+		/* xchacha with 192-bit iv */
+		u32int counter[2];
+		uchar h[32];
+
+		s->input[4] = s->xkey[0];
+		s->input[5] = s->xkey[1];
+		s->input[6] = s->xkey[2];
+		s->input[7] = s->xkey[3];
+		s->input[8] = s->xkey[4];
+		s->input[9] = s->xkey[5];
+		s->input[10] = s->xkey[6];
+		s->input[11] = s->xkey[7];
+
+		counter[0] = s->input[12];
+		counter[1] = s->input[13];
+
+		load(&s->input[12], iv, 4);
+
+		hchachablock(h, s);
+		load(&s->input[4], h, 8);
+		memset(h, 0, 32);
+
+		s->input[12] = counter[0];
+		s->input[13] = counter[1];
+
+		load(&s->input[14], iv+16, 2);
+		return;
+	}
 	load(&s->input[16 - s->ivwords], iv, s->ivwords);
 }
 
@@ -90,7 +176,7 @@
 chacha_setblock(Chachastate *s, u64int blockno)
 {
 	s->input[12] = blockno;
-	if(s->ivwords == 2)
+	if(s->ivwords != 3)
 		s->input[13] = blockno>>32;
 }
 
@@ -98,9 +184,8 @@
 encryptblock(Chachastate *s, uchar *src, uchar *dst)
 {
 	u32int x[Blockwords];
-	int i, rounds;
+	int i;
 
-	rounds = s->rounds;
 	x[0] = s->input[0];
 	x[1] = s->input[1];
 	x[2] = s->input[2];
@@ -117,37 +202,8 @@
 	x[13] = s->input[13];
 	x[14] = s->input[14];
 	x[15] = s->input[15];
+	dorounds(x, s->rounds);
 
-	for(i = rounds; i > 0; i -= 2) {
-		QUARTERROUND(0, 4, 8,12)
-		QUARTERROUND(1, 5, 9,13)
-		QUARTERROUND(2, 6,10,14)
-		QUARTERROUND(3, 7,11,15)
-
-		QUARTERROUND(0, 5,10,15)
-		QUARTERROUND(1, 6,11,12)
-		QUARTERROUND(2, 7, 8,13)
-		QUARTERROUND(3, 4, 9,14)
-	}
-
-#ifdef FULL_UNROLL
-	ENCRYPT(src+0*4, x[0], s->input[0], dst+0*4);
-	ENCRYPT(src+1*4, x[1], s->input[1], dst+1*4);
-	ENCRYPT(src+2*4, x[2], s->input[2], dst+2*4);
-	ENCRYPT(src+3*4, x[3], s->input[3], dst+3*4);
-	ENCRYPT(src+4*4, x[4], s->input[4], dst+4*4);
-	ENCRYPT(src+5*4, x[5], s->input[5], dst+5*4);
-	ENCRYPT(src+6*4, x[6], s->input[6], dst+6*4);
-	ENCRYPT(src+7*4, x[7], s->input[7], dst+7*4);
-	ENCRYPT(src+8*4, x[8], s->input[8], dst+8*4);
-	ENCRYPT(src+9*4, x[9], s->input[9], dst+9*4);
-	ENCRYPT(src+10*4, x[10], s->input[10], dst+10*4);
-	ENCRYPT(src+11*4, x[11], s->input[11], dst+11*4);
-	ENCRYPT(src+12*4, x[12], s->input[12], dst+12*4);
-	ENCRYPT(src+13*4, x[13], s->input[13], dst+13*4);
-	ENCRYPT(src+14*4, x[14], s->input[14], dst+14*4);
-	ENCRYPT(src+15*4, x[15], s->input[15], dst+15*4);
-#else
 	for(i=0; i<nelem(x); i+=4){
 		ENCRYPT(src, x[i], s->input[i], dst);
 		ENCRYPT(src+4, x[i+1], s->input[i+1], dst+4);
@@ -156,9 +212,8 @@
 		src += 16;
 		dst += 16;
 	}
-#endif
 
-	if(++s->input[12] == 0 && s->ivwords == 2)
+	if(++s->input[12] == 0 && s->ivwords != 3)
 		s->input[13]++;
 }
 
@@ -183,4 +238,14 @@
 chacha_encrypt(uchar *buf, ulong bytes, Chachastate *s)
 {
 	chacha_encrypt2(buf, buf, bytes, s);
+}
+
+void
+hchacha(uchar h[32], uchar *key, ulong keylen, uchar nonce[16], int rounds)
+{
+	Chachastate s[1];
+
+	setupChachastate(s, key, keylen, nonce, 16, rounds);
+	hchachablock(h, s);
+	memset(s, 0, sizeof(s));
 }
--- a/libsec/salsa.c
+++ b/libsec/salsa.c
@@ -1,24 +1,17 @@
 #include "os.h"
 #include <libsec.h>
 
-enum{
-	Blockwords=	SalsaBsize/sizeof(u32int)
-};
-
 /* little-endian data order */
-#define GET4(p)	((((((p)[3]<<8) | (p)[2])<<8) | (p)[1])<<8 | (p)[0])
-#define PUT4(p, v)	(((p)[0]=v), (v>>=8), ((p)[1]=v), (v>>=8), ((p)[2]=v), (v>>=8), ((p)[3]=v))
+#define	GET4(p)		((p)[0]|((p)[1]<<8)|((p)[2]<<16)|((p)[3]<<24))
+#define	PUT4(p,v)	(p)[0]=(v);(p)[1]=(v)>>8;(p)[2]=(v)>>16;(p)[3]=(v)>>24
 
 #define ROTATE(v,c) (t = v, (u32int)(t << (c)) | (t >> (32 - (c))))
 
 #define ENCRYPT(s, x, y, d) {\
 	u32int v; \
-	uchar *sp, *dp; \
-	sp = (s); \
-	v = GET4(sp); \
+	v = GET4(s); \
 	v ^= (x)+(y); \
-	dp = (d); \
-	PUT4(dp, v); \
+	PUT4(d, v); \
 }
 
 static uchar sigma[16] = "expand 32-byte k";
@@ -38,7 +31,8 @@
 {
 	if(keylen != 256/8 && keylen != 128/8)
 		sysfatal("invalid salsa key length");
-	if(ivlen != 64/8 && ivlen != 128/8 && ivlen != 192/8)
+	if(ivlen != 64/8
+	&& ivlen != 128/8 && ivlen != 192/8)	/* hsalsa, xsalsa */
 		sysfatal("invalid salsa iv length");
 	if(rounds == 0)
 		rounds = 20;
@@ -58,14 +52,14 @@
 		load(&s->input[11], key, 4);
 		load(&s->input[15], tau +4*3, 1);
 	}
-	s->key[0] = s->input[1];
-	s->key[1] = s->input[2];
-	s->key[2] = s->input[3];
-	s->key[3] = s->input[4];
-	s->key[4] = s->input[11];
-	s->key[5] = s->input[12];
-	s->key[6] = s->input[13];
-	s->key[7] = s->input[14];
+	s->xkey[0] = s->input[1];
+	s->xkey[1] = s->input[2];
+	s->xkey[2] = s->input[3];
+	s->xkey[3] = s->input[4];
+	s->xkey[4] = s->input[11];
+	s->xkey[5] = s->input[12];
+	s->xkey[6] = s->input[13];
+	s->xkey[7] = s->input[14];
 
 	s->ivwords = ivlen/4;
 	s->input[8] = 0;
@@ -78,30 +72,11 @@
 }
 
 static void
-hsalsablock(uchar h[32], Salsastate *s)
+dorounds(u32int x[16], int rounds)
 {
-	u32int x[Blockwords], t;
-	int i, rounds;
+	u32int t;
 
-	rounds = s->rounds;
-	x[0] = s->input[0];
-	x[1] = s->input[1];
-	x[2] = s->input[2];
-	x[3] = s->input[3];
-	x[4] = s->input[4];
-	x[5] = s->input[5];
-	x[6] = s->input[6];
-	x[7] = s->input[7];
-	x[8] = s->input[8];
-	x[9] = s->input[9];
-	x[10] = s->input[10];
-	x[11] = s->input[11];
-	x[12] = s->input[12];
-	x[13] = s->input[13];
-	x[14] = s->input[14];
-	x[15] = s->input[15];
-
-	for(i = rounds; i > 0; i -= 2) {
+	for(; rounds > 0; rounds -= 2) {
 	     x[4] ^= ROTATE( x[0]+x[12], 7);
 	     x[8] ^= ROTATE( x[4]+ x[0], 9);
 	    x[12] ^= ROTATE( x[8]+ x[4],13);
@@ -135,7 +110,32 @@
 	    x[14] ^= ROTATE(x[13]+x[12],13);
 	    x[15] ^= ROTATE(x[14]+x[13],18);
 	}
+}
 
+static void
+hsalsablock(uchar h[32], Salsastate *s)
+{
+	u32int x[16];
+
+	x[0] = s->input[0];
+	x[1] = s->input[1];
+	x[2] = s->input[2];
+	x[3] = s->input[3];
+	x[4] = s->input[4];
+	x[5] = s->input[5];
+	x[6] = s->input[6];
+	x[7] = s->input[7];
+	x[8] = s->input[8];
+	x[9] = s->input[9];
+	x[10] = s->input[10];
+	x[11] = s->input[11];
+	x[12] = s->input[12];
+	x[13] = s->input[13];
+	x[14] = s->input[14];
+	x[15] = s->input[15];
+
+	dorounds(x, s->rounds);
+
 	PUT4(h+0*4, x[0]);
 	PUT4(h+1*4, x[5]);
 	PUT4(h+2*4, x[10]);
@@ -150,7 +150,7 @@
 salsa_setiv(Salsastate *s, uchar *iv)
 {
 	if(s->ivwords == 128/32){
-		/* hsalsa 128-bit iv */
+		/* hsalsa with 128-bit iv */
 		load(&s->input[6], iv, 4);
 		return;
 	}
@@ -162,14 +162,14 @@
 		counter[0] = s->input[8];
 		counter[1] = s->input[9];
 
-		s->input[1] = s->key[0];
-		s->input[2] = s->key[1];
-		s->input[3] = s->key[2];
-		s->input[4] = s->key[3];
-		s->input[11] = s->key[4];
-		s->input[12] = s->key[5];
-		s->input[13] = s->key[6];
-		s->input[14] = s->key[7];
+		s->input[1] = s->xkey[0];
+		s->input[2] = s->xkey[1];
+		s->input[3] = s->xkey[2];
+		s->input[4] = s->xkey[3];
+		s->input[11] = s->xkey[4];
+		s->input[12] = s->xkey[5];
+		s->input[13] = s->xkey[6];
+		s->input[14] = s->xkey[7];
 
 		load(&s->input[6], iv, 4);
 
@@ -197,10 +197,9 @@
 static void
 encryptblock(Salsastate *s, uchar *src, uchar *dst)
 {
-	u32int x[Blockwords], t;
-	int i, rounds;
+	u32int x[16];
+	int i;
 
-	rounds = s->rounds;
 	x[0] = s->input[0];
 	x[1] = s->input[1];
 	x[2] = s->input[2];
@@ -218,59 +217,8 @@
 	x[14] = s->input[14];
 	x[15] = s->input[15];
 
-	for(i = rounds; i > 0; i -= 2) {
-	     x[4] ^= ROTATE( x[0]+x[12], 7);
-	     x[8] ^= ROTATE( x[4]+ x[0], 9);
-	    x[12] ^= ROTATE( x[8]+ x[4],13);
-	     x[0] ^= ROTATE(x[12]+ x[8],18);
-	     x[9] ^= ROTATE( x[5]+ x[1], 7);
-	    x[13] ^= ROTATE( x[9]+ x[5], 9);
-	     x[1] ^= ROTATE(x[13]+ x[9],13);
-	     x[5] ^= ROTATE( x[1]+x[13],18);
-	    x[14] ^= ROTATE(x[10]+ x[6], 7);
-	     x[2] ^= ROTATE(x[14]+x[10], 9);
-	     x[6] ^= ROTATE( x[2]+x[14],13);
-	    x[10] ^= ROTATE( x[6]+ x[2],18);
-	     x[3] ^= ROTATE(x[15]+x[11], 7);
-	     x[7] ^= ROTATE( x[3]+x[15], 9);
-	    x[11] ^= ROTATE( x[7]+ x[3],13);
-	    x[15] ^= ROTATE(x[11]+ x[7],18);
-	     x[1] ^= ROTATE( x[0]+ x[3], 7);
-	     x[2] ^= ROTATE( x[1]+ x[0], 9);
-	     x[3] ^= ROTATE( x[2]+ x[1],13);
-	     x[0] ^= ROTATE( x[3]+ x[2],18);
-	     x[6] ^= ROTATE( x[5]+ x[4], 7);
-	     x[7] ^= ROTATE( x[6]+ x[5], 9);
-	     x[4] ^= ROTATE( x[7]+ x[6],13);
-	     x[5] ^= ROTATE( x[4]+ x[7],18);
-	    x[11] ^= ROTATE(x[10]+ x[9], 7);
-	     x[8] ^= ROTATE(x[11]+x[10], 9);
-	     x[9] ^= ROTATE( x[8]+x[11],13);
-	    x[10] ^= ROTATE( x[9]+ x[8],18);
-	    x[12] ^= ROTATE(x[15]+x[14], 7);
-	    x[13] ^= ROTATE(x[12]+x[15], 9);
-	    x[14] ^= ROTATE(x[13]+x[12],13);
-	    x[15] ^= ROTATE(x[14]+x[13],18);
-	}
+	dorounds(x, s->rounds);
 
-#ifdef FULL_UNROLL
-	ENCRYPT(src+0*4, x[0], s->input[0], dst+0*4);
-	ENCRYPT(src+1*4, x[1], s->input[1], dst+1*4);
-	ENCRYPT(src+2*4, x[2], s->input[2], dst+2*4);
-	ENCRYPT(src+3*4, x[3], s->input[3], dst+3*4);
-	ENCRYPT(src+4*4, x[4], s->input[4], dst+4*4);
-	ENCRYPT(src+5*4, x[5], s->input[5], dst+5*4);
-	ENCRYPT(src+6*4, x[6], s->input[6], dst+6*4);
-	ENCRYPT(src+7*4, x[7], s->input[7], dst+7*4);
-	ENCRYPT(src+8*4, x[8], s->input[8], dst+8*4);
-	ENCRYPT(src+9*4, x[9], s->input[9], dst+9*4);
-	ENCRYPT(src+10*4, x[10], s->input[10], dst+10*4);
-	ENCRYPT(src+11*4, x[11], s->input[11], dst+11*4);
-	ENCRYPT(src+12*4, x[12], s->input[12], dst+12*4);
-	ENCRYPT(src+13*4, x[13], s->input[13], dst+13*4);
-	ENCRYPT(src+14*4, x[14], s->input[14], dst+14*4);
-	ENCRYPT(src+15*4, x[15], s->input[15], dst+15*4);
-#else
 	for(i=0; i<nelem(x); i+=4){
 		ENCRYPT(src, x[i], s->input[i], dst);
 		ENCRYPT(src+4, x[i+1], s->input[i+1], dst+4);
@@ -279,7 +227,6 @@
 		src += 16;
 		dst += 16;
 	}
-#endif
 
 	if(++s->input[8] == 0)
 		s->input[9]++;
@@ -306,6 +253,48 @@
 salsa_encrypt(uchar *buf, ulong bytes, Salsastate *s)
 {
 	salsa_encrypt2(buf, buf, bytes, s);
+}
+
+void
+salsa_core(u32int in[16], u32int out[16], int rounds)
+{
+	u32int x[16];
+
+	x[0] = in[0];
+	x[1] = in[1];
+	x[2] = in[2];
+	x[3] = in[3];
+	x[4] = in[4];
+	x[5] = in[5];
+	x[6] = in[6];
+	x[7] = in[7];
+	x[8] = in[8];
+	x[9] = in[9];
+	x[10] = in[10];
+	x[11] = in[11];
+	x[12] = in[12];
+	x[13] = in[13];
+	x[14] = in[14];
+	x[15] = in[15];
+
+	dorounds(x, rounds);
+
+	out[0] = x[0] + in[0];
+	out[1] = x[1] + in[1];
+	out[2] = x[2] + in[2];
+	out[3] = x[3] + in[3];
+	out[4] = x[4] + in[4];
+	out[5] = x[5] + in[5];
+	out[6] = x[6] + in[6];
+	out[7] = x[7] + in[7];
+	out[8] = x[8] + in[8];
+	out[9] = x[9] + in[9];
+	out[10] = x[10] + in[10];
+	out[11] = x[11] + in[11];
+	out[12] = x[12] + in[12];
+	out[13] = x[13] + in[13];
+	out[14] = x[14] + in[14];
+	out[15] = x[15] + in[15];
 }
 
 void
--- /dev/null
+++ b/libsec/scrypt.c
@@ -1,0 +1,119 @@
+#include "os.h"
+#include <libsec.h>
+
+#define movw(w, S, D)	memmove(D, S, (w)*4)
+
+static void
+xorw(ulong w, u32int *S, u32int *D)
+{
+	for(w /= 8; w; w--, D += 8, S += 8){
+		D[0] ^= S[0];
+		D[1] ^= S[1];
+		D[2] ^= S[2];
+		D[3] ^= S[3];
+		D[4] ^= S[4];
+		D[5] ^= S[5];
+		D[6] ^= S[6];
+		D[7] ^= S[7];
+	}
+}
+
+static void
+scryptBlockMix(ulong R, u32int *B, u32int *Y)
+{
+	u32int X[16];
+	ulong i;
+
+	R *= 2;
+	movw(16, &B[(R-1)*16], X);
+	for(i = 0; i < R; i += 2){
+		xorw(16, &B[i*16], X);
+		salsa_core(X, X, 8);
+		movw(16, X, &Y[i*8]);
+
+		xorw(16, &B[(i+1)*16], X);
+		salsa_core(X, X, 8);
+		movw(16, X, &Y[i*8 + R*8]);
+	}
+}
+
+static void
+scryptROMix(ulong R, ulong N, u32int *V, u32int *X, uchar *B)
+{
+	ulong w, i, d;
+	u32int *Y;
+
+	w = R*32;
+	for(i=0; i<w; i++, B+=4)
+		X[i] = B[0] | (B[1]<<8) | (B[2]<<16) | (B[3]<<24);
+
+	Y = &X[w];
+	for(i=0; i<N; i += 2){
+		movw(w, X, &V[i*w]);
+		scryptBlockMix(R, X, Y);
+
+		movw(w, Y, &V[(i+1)*w]);
+		scryptBlockMix(R, Y, X);
+	}
+	for(i=0; i<N; i += 2){
+		xorw(w, &V[(X[w-16] & (N-1))*w], X);
+		scryptBlockMix(R, X, Y);
+
+		xorw(w, &V[(Y[w-16] & (N-1))*w], Y);
+		scryptBlockMix(R, Y, X);
+	}
+
+	B -= w*4;
+	for(i=0; i<w; i++, B+=4)
+		d = X[i], B[0]=d, B[1]=d>>8, B[2]=d>>16, B[3]=d>>24;
+}
+
+char*
+scrypt(p, plen, s, slen, N, R, P, d, dlen)
+	ulong plen, slen, dlen, N, R, P;
+	uchar *p, *s, *d;
+{
+	static char oom[] = "out of memory";
+
+	ulong rb, i;
+	u32int *V, *X;
+	uchar *B;
+
+	if(P < 1)
+		return "invalid parallelization parameter P";
+	if(R < 1 || R >= (1UL<<(31-7))/P)
+		return "invalid block size parameter R";
+	if(N < 2 || (N & (N-1)) != 0 || N >= (1UL<<(31-7))/R)
+		return "invalid cpu/memory cost parameter N";
+
+	rb = R<<7;
+	if((B = malloc(P*rb)) == nil)
+		return oom;
+	if((V = malloc(N*rb)) == nil){
+		free(B);
+		return oom;
+	}
+	if((X = malloc(2*rb)) == nil){
+		free(V);
+		free(B);
+		return oom;
+	}
+
+	pbkdf2_x(p, plen, s, slen, 1, B, P*rb, hmac_sha2_256, SHA2_256dlen);
+
+	for(i=0; i<P; i++)
+		scryptROMix(R, N, V, X, &B[i*rb]);
+
+	memset(X, 0, 2*rb);
+	free(X);
+
+	memset(V, 0, N*rb);
+	free(V);
+
+	pbkdf2_x(p, plen, B, P*rb, 1, d, dlen, hmac_sha2_256, SHA2_256dlen);
+
+	memset(B, 0, P*rb);
+	free(B);
+
+	return nil;
+}