X-Git-Url: https://git.distorted.org.uk/~mdw/catacomb/blobdiff_plain/2b3f6527d522b647a6c8b5039228ebe569686c38..HEAD:/symm/salsa20.c

diff --git a/symm/salsa20.c b/symm/salsa20.c
index f424b746..f0fe3d7c 100644
--- a/symm/salsa20.c
+++ b/symm/salsa20.c
@@ -39,6 +39,7 @@
 #include "grand.h"
 #include "keysz.h"
 #include "paranoia.h"
+#include "rsvr.h"
 #include "salsa20.h"
 #include "salsa20-core.h"
 
@@ -72,6 +73,7 @@ static void simple_core(unsigned r, const salsa20_matrix src,
 
 #if CPUFAM_X86 || CPUFAM_AMD64
 extern core__functype salsa20_core_x86ish_sse2;
+extern core__functype salsa20_core_x86ish_avx;
 #endif
 
 #if CPUFAM_ARMEL
@@ -85,6 +87,8 @@ extern core__functype salsa20_core_arm64;
 static core__functype *pick_core(void)
 {
 #if CPUFAM_X86 || CPUFAM_AMD64
+  DISPATCH_PICK_COND(salsa20_core, salsa20_core_x86ish_avx,
+		     cpu_feature_p(CPUFEAT_X86_AVX));
   DISPATCH_PICK_COND(salsa20_core, salsa20_core_x86ish_sse2,
 		     cpu_feature_p(CPUFEAT_X86_SSE2));
 #endif
@@ -159,6 +163,8 @@ static void populate(salsa20_matrix a, const void *key, size_t ksz)
 
 /*----- Salsa20 implementation --------------------------------------------*/
 
+static const octet zerononce[XSALSA20_NONCESZ];
+
 /* --- @salsa20_init@ --- *
  *
  * Arguments:	@salsa20_ctx *ctx@ = context to fill in
@@ -174,8 +180,6 @@ static void populate(salsa20_matrix a, const void *key, size_t ksz)
 void salsa20_init(salsa20_ctx *ctx, const void *key, size_t ksz,
 		  const void *nonce)
 {
-  static const octet zerononce[SALSA20_NONCESZ];
-
   populate(ctx->a, key, ksz);
   salsa20_setnonce(ctx, nonce ? nonce : zerononce);
 }
@@ -231,7 +235,7 @@ void salsa20_seek(salsa20_ctx *ctx, unsigned long i)
 void salsa20_seeku64(salsa20_ctx *ctx, kludge64 i)
 {
   ctx->a[8] = LO64(i); ctx->a[5] = HI64(i);
-  ctx->bufi = SALSA20_OUTSZ;
+  ctx->off = 0;
 }
 
 void salsa20_seek_ietf(salsa20_ctx *ctx, uint32 i)
@@ -271,6 +275,8 @@ uint32 salsa20_tell_ietf(salsa20_ctx *ctx)
  *		to @dest@.
  */
 
+static const rsvr_policy policy = { 0, SALSA20_OUTSZ, SALSA20_OUTSZ };
+
 #define SALSA20_ENCRYPT(r, ctx, src, dest, sz)				\
   SALSA20_DECOR(salsa20, r, _encrypt)(ctx, src, dest, sz)
 #define DEFENCRYPT(r)							\
@@ -280,41 +286,40 @@ uint32 salsa20_tell_ietf(salsa20_ctx *ctx)
     salsa20_matrix b;							\
     const octet *s = src;						\
     octet *d = dest;							\
-    size_t n;								\
+    rsvr_plan plan;							\
     kludge64 pos, delta;						\
 									\
-    SALSA20_OUTBUF(ctx, d, s, sz);					\
-    if (!sz) return;							\
-									\
-    if (!dest) {							\
-      n = sz/SALSA20_OUTSZ;						\
-      pos = salsa20_tellu64(ctx);					\
-      ASSIGN64(delta, n);						\
-      ADD64(pos, pos, delta);						\
-      salsa20_seeku64(ctx, pos);					\
-      sz = sz%SALSA20_OUTSZ;						\
-    } else if (!src) {							\
-      while (sz >= SALSA20_OUTSZ) {					\
-	core(r, ctx->a, b);						\
-	SALSA20_STEP(ctx->a);						\
-	SALSA20_GENFULL(b, d);						\
-	sz -= SALSA20_OUTSZ;						\
+    rsvr_mkplan(&plan, &policy, ctx->off, sz);				\
+									\
+    if (plan.head) {							\
+      if (!ctx->off) {							\
+	core(r, ctx->a, b); SALSA20_STEP(ctx->a);			\
+	SALSA20_PREPBUF(ctx, b);					\
       }									\
-    } else {								\
-      while (sz >= SALSA20_OUTSZ) {					\
-	core(r, ctx->a, b);						\
-	SALSA20_STEP(ctx->a);						\
-	SALSA20_MIXFULL(b, d, s);					\
-	sz -= SALSA20_OUTSZ;						\
+      SALSA20_OUTBUF(ctx, d, s, plan.head);				\
+    }									\
+									\
+    ctx->off -= plan.from_rsvr;						\
+									\
+    if (!d) {								\
+      if (plan.from_input) {						\
+	pos = salsa20_tellu64(ctx);					\
+	ASSIGN64(delta, plan.from_input/SALSA20_OUTSZ);			\
+	ADD64(pos, pos, delta);						\
+	salsa20_seeku64(ctx, pos);					\
       }									\
+    } else if (!s) while (plan.from_input) {				\
+      core(r, ctx->a, b); SALSA20_STEP(ctx->a);				\
+      SALSA20_GENFULL(b, d); plan.from_input -= SALSA20_OUTSZ;		\
+    } else while (plan.from_input) {					\
+      core(r, ctx->a, b); SALSA20_STEP(ctx->a);				\
+      SALSA20_MIXFULL(b, d, s); plan.from_input -= SALSA20_OUTSZ;	\
     }									\
 									\
-    if (sz) {								\
-      core(r, ctx->a, b);						\
-      SALSA20_STEP(ctx->a);						\
+    if (plan.tail) {							\
+      core(r, ctx->a, b); SALSA20_STEP(ctx->a);				\
       SALSA20_PREPBUF(ctx, b);						\
-      SALSA20_OUTBUF(ctx, d, s, sz);					\
-      assert(!sz);							\
+      SALSA20_OUTBUF(ctx, d, s, plan.tail);				\
     }									\
   }
 SALSA20_VARS(DEFENCRYPT)
@@ -419,8 +424,6 @@ SALSA20_VARS(DEFHSALSA20)
   void XSALSA20_INIT(r, XSALSA20_CTX(r) *ctx,				\
 			const void *key, size_t ksz, const void *nonce)	\
   {									\
-    static const octet zerononce[XSALSA20_NONCESZ];			\
-									\
     populate(ctx->k, key, ksz);						\
     ctx->s.a[ 0] = SALSA20_A256;					\
     ctx->s.a[ 1] = SALSA20_B256;					\
@@ -868,9 +871,14 @@ SALSA20_VARS(DEFXGRAND)
 #include <stdio.h>
 #include <string.h>
 
+#include <mLib/macros.h>
 #include <mLib/quis.h>
 #include <mLib/testrig.h>
 
+#ifdef ENABLE_ASM_DEBUG
+#  include "regdump.h"
+#endif
+
 static const int perm[] = {
    0, 13, 10,  7,
    4,  1, 14, 11,
@@ -897,7 +905,7 @@ static const int perm[] = {
     }									\
     for (i = 0; i < SALSA20_OUTSZ/4; i++) STORE32_L(d.buf + 4*i, b[i]);	\
 									\
-    if (d.len != v[2].len || memcmp(d.buf, v[2].buf, v[2].len) != 0) {	\
+    if (d.len != v[2].len || MEMCMP(d.buf, !=, v[2].buf, v[2].len)) {	\
       ok = 0;								\
       printf("\nfail core:"						\
 	     "\n\titerations = %d"					\
@@ -967,7 +975,7 @@ SALSA20_VARS(DEFVCORE)
       }									\
       if (sz) BASE##_ENCRYPT(r, &ctx, p, q, sz);			\
 									\
-      if (d.len != v[5].len || memcmp(d.buf, v[5].buf, v[5].len) != 0) { \
+      if (d.len != v[5].len || MEMCMP(d.buf, !=, v[5].buf, v[5].len)) {	\
 	ok = 0;								\
 	printf("\nfail encrypt:"					\
 	       "\n\tstep	   = %lu"				\
@@ -1013,6 +1021,9 @@ SALSA20_VARS(DEFXTAB)
 
 int main(int argc, char *argv[])
 {
+#ifdef ENABLE_ASM_DEBUG
+  regdump_init();
+#endif
   test_run(argc, argv, defs, SRCDIR"/t/salsa20");
   return (0);
 }