From 86abac78a20dfe49aea28058956a3fe8b6ced9ee Mon Sep 17 00:00:00 2001
From: Werner Koch <wk@gnupg.org>
Date: Fri, 2 Jul 1999 09:50:57 +0000
Subject: See ChangeLog: Fri Jul  2 11:45:54 CEST 1999  Werner Koch

---
 mpi/ChangeLog      |   6 +++
 mpi/mpi-bit.c      |  48 ++++++++++++++++++
 mpi/mpi-internal.h |   5 ++
 mpi/mpi-mpow.c     | 141 +++++++++++++++++++++++++++++++++++++++++++++--------
 mpi/mpi-mul.c      |   1 -
 5 files changed, 181 insertions(+), 20 deletions(-)

(limited to 'mpi')

diff --git a/mpi/ChangeLog b/mpi/ChangeLog
index 959f3b48c..1951dd3a4 100644
--- a/mpi/ChangeLog
+++ b/mpi/ChangeLog
@@ -1,3 +1,9 @@
+Fri Jul  2 11:45:54 CEST 1999  Werner Koch  <wk@isil.d.shuttle.de>
+
+
+	* mpi-bit.c (mpi_lshift_limbs,mpi_rshift_limbs): New.
+	* mpi-mpow.c (barrett_mulm): New but diabled.
+
 Tue Jun  1 16:01:46 CEST 1999  Werner Koch  <wk@isil.d.shuttle.de>
 
 	* config.links (i[56]86*-*-freebsdelf*): New.
diff --git a/mpi/mpi-bit.c b/mpi/mpi-bit.c
index 00aa5d086..f1eff8636 100644
--- a/mpi/mpi-bit.c
+++ b/mpi/mpi-bit.c
@@ -212,3 +212,51 @@ mpi_rshift( MPI x, MPI a, unsigned n )
     x->nlimbs = xsize;
 }
 
+
+/****************
+ * Shift A by COUNT limbs to the left
+ * This is used only within the MPI library
+ */
+void
+mpi_lshift_limbs( MPI a, unsigned int count )
+{
+    mpi_ptr_t ap = a->d;
+    int n = a->nlimbs;
+    int i;
+
+    if( !count || !n )
+	return;
+
+    RESIZE_IF_NEEDED( a, n+count );
+
+    for( i = n-1; i >= 0; i-- )
+	ap[i+count] = ap[i];
+    for(i=0; i < count; i++ )
+	ap[i] = 0;
+    a->nlimbs += count;
+}
+
+
+/****************
+ * Shift A by COUNT limbs to the right
+ * This is used only within the MPI library
+ */
+void
+mpi_rshift_limbs( MPI a, unsigned int count )
+{
+    mpi_ptr_t ap = a->d;
+    mpi_size_t n = a->nlimbs;
+    unsigned int i;
+
+    if( count >= n ) {
+	a->nlimbs = 0;
+	return;
+    }
+
+    for( i = 0; i < n - count; i++ )
+	ap[i] = ap[i+count];
+    ap[i] = 0;
+    a->nlimbs -= count;
+}
+
+
diff --git a/mpi/mpi-internal.h b/mpi/mpi-internal.h
index f73efb76c..035d33cb3 100644
--- a/mpi/mpi-internal.h
+++ b/mpi/mpi-internal.h
@@ -161,6 +161,11 @@ typedef int mpi_size_t;        /* (must be a signed type) */
 #endif
 void mpi_assign_limb_space( MPI a, mpi_ptr_t ap, unsigned nlimbs );
 
+/*-- mpi-bit.c --*/
+void mpi_rshift_limbs( MPI a, unsigned int count );
+void mpi_lshift_limbs( MPI a, unsigned int count );
+
+
 /*-- mpihelp-add.c --*/
 mpi_limb_t mpihelp_add_1(mpi_ptr_t res_ptr,  mpi_ptr_t s1_ptr,
 			 mpi_size_t s1_size, mpi_limb_t s2_limb );
diff --git a/mpi/mpi-mpow.c b/mpi/mpi-mpow.c
index 689a7600f..a8c561dd1 100644
--- a/mpi/mpi-mpow.c
+++ b/mpi/mpi-mpow.c
@@ -1,5 +1,5 @@
 /* mpi-mpow.c  -  MPI functions
- *	Copyright (C) 1998 Free Software Foundation, Inc.
+ *	Copyright (C) 1998, 1999 Free Software Foundation, Inc.
  *
  * This file is part of GnuPG.
  *
@@ -25,6 +25,22 @@
 #include "longlong.h"
 #include <assert.h>
 
+/* Barrett is slower than the classical way.  It can be tweaked by
+ * using partial multiplications
+ */
+/*#define USE_BARRETT*/
+
+
+
+#ifdef USE_BARRETT
+static void barrett_mulm( MPI w, MPI u, MPI v, MPI m, MPI y, int k, MPI r1, MPI r2 );
+static MPI init_barrett( MPI m, int *k, MPI *r1, MPI *r2 );
+static int calc_barrett( MPI r, MPI x, MPI m, MPI y, int k, MPI r1, MPI r2  );
+#else
+#define barrett_mulm( w, u, v, m, y, k, r1, r2 ) mpi_mulm( (w), (u), (v), (m) )
+#endif
+
+
 static int
 build_index( MPI *exparray, int k, int i, int t )
 {
@@ -52,6 +68,10 @@ mpi_mulpowm( MPI res, MPI *basearray, MPI *exparray, MPI m)
     int i, j, idx;
     MPI *G;	/* table with precomputed values of size 2^k */
     MPI tmp;
+  #ifdef USE_BARRETT
+    MPI barrett_y, barrett_r1, barrett_r2;
+    int barrett_k;
+  #endif
 
     for(k=0; basearray[k]; k++ )
 	;
@@ -68,27 +88,15 @@ mpi_mulpowm( MPI res, MPI *basearray, MPI *exparray, MPI m)
     assert( k < 10 );
 
     G = m_alloc_clear( (1<<k) * sizeof *G );
-  #if 0
-    /* do the precomputation */
-    G[0] = mpi_alloc_set_ui( 1 );
-    for(i=1; i < (1<<k); i++ ) {
-	for(j=0; j < k; j++ ) {
-	    if( (i & (1<<j) ) ) {
-		if( !G[i] )
-		    G[i] = mpi_copy( basearray[j] );
-		else
-		    mpi_mulm( G[i], G[i], basearray[j], m );
-	    }
-	}
-	if( !G[i] )
-	    G[i] = mpi_alloc(0);
-    }
+  #ifdef USE_BARRETT
+    barrett_y = init_barrett( m, &barrett_k, &barrett_r1, &barrett_r2 );
   #endif
     /* and calculate */
     tmp =  mpi_alloc( mpi_get_nlimbs(m)+1 );
     mpi_set_ui( res, 1 );
     for(i = 1; i <= t; i++ ) {
-	mpi_mulm(tmp, res, res, m );
+	barrett_mulm(tmp, res, res, m, barrett_y, barrett_k,
+				       barrett_r1, barrett_r2 );
 	idx = build_index( exparray, k, i, t );
 	assert( idx >= 0 && idx < (1<<k) );
 	if( !G[idx] ) {
@@ -100,20 +108,115 @@ mpi_mulpowm( MPI res, MPI *basearray, MPI *exparray, MPI m)
 			if( !G[idx] )
 			    G[idx] = mpi_copy( basearray[j] );
 			else
-			    mpi_mulm( G[idx], G[idx], basearray[j], m );
+			    barrett_mulm( G[idx], G[idx], basearray[j],
+					       m, barrett_y, barrett_k, barrett_r1, barrett_r2	);
 		    }
 		}
 		if( !G[idx] )
 		    G[idx] = mpi_alloc(0);
 	    }
 	}
-	mpi_mulm(res, tmp, G[idx], m );
+	barrett_mulm(res, tmp, G[idx], m, barrett_y, barrett_k, barrett_r1, barrett_r2	);
     }
 
     /* cleanup */
     mpi_free(tmp);
+  #ifdef USE_BARRETT
+    mpi_free(barrett_y);
+    mpi_free(barrett_r1);
+    mpi_free(barrett_r2);
+  #endif
     for(i=0; i < (1<<k); i++ )
 	mpi_free(G[i]);
     m_free(G);
 }
 
+
+
+#ifdef USE_BARRETT
+static void
+barrett_mulm( MPI w, MPI u, MPI v, MPI m, MPI y, int k, MPI r1, MPI r2	)
+{
+    mpi_mul(w, u, v);
+    if( calc_barrett( w, w, m, y, k, r1, r2 ) )
+	mpi_fdiv_r( w, w, m );
+}
+
+/****************
+ * Barrett precalculation: y = floor(b^(2k) / m)
+ */
+static MPI
+init_barrett( MPI m, int *k, MPI *r1, MPI *r2 )
+{
+    MPI tmp;
+
+    mpi_normalize( m );
+    *k = mpi_get_nlimbs( m );
+    tmp = mpi_alloc( *k + 1 );
+    mpi_set_ui( tmp, 1 );
+    mpi_lshift_limbs( tmp, 2 * *k );
+    mpi_fdiv_q( tmp, tmp, m );
+    *r1 = mpi_alloc( 2* *k + 1 );
+    *r2 = mpi_alloc( 2* *k + 1 );
+    return tmp;
+}
+
+/****************
+ * Barrett reduction: We assume that these conditions are met:
+ * Given x =(x_2k-1 ...x_0)_b
+ *	 m =(m_k-1 ....m_0)_b	  with m_k-1 != 0
+ * Output r = x mod m
+ * Before using this function init_barret must be used to calucalte y and k.
+ * Returns: false = no error
+ *	    true = can't perform barret reduction
+ */
+static int
+calc_barrett( MPI r, MPI x, MPI m, MPI y, int k, MPI r1, MPI r2 )
+{
+    int xx = k > 3 ? k-3:0;
+
+    mpi_normalize( x );
+    if( mpi_get_nlimbs(x) > 2*k )
+	return 1; /* can't do it */
+
+    /* 1. q1 = floor( x / b^k-1)
+     *	  q2 = q1 * y
+     *	  q3 = floor( q2 / b^k+1 )
+     * Actually, we don't need qx, we can work direct on r2
+     */
+    mpi_set( r2, x );
+    mpi_rshift_limbs( r2, k-1 );
+    mpi_mul( r2, r2, y );
+    mpi_rshift_limbs( r2, k+1 );
+
+    /* 2. r1 = x mod b^k+1
+     *	  r2 = q3 * m mod b^k+1
+     *	  r  = r1 - r2
+     * 3. if r < 0 then  r = r + b^k+1
+     */
+    mpi_set( r1, x );
+    if( r1->nlimbs > k+1 ) /* quick modulo operation */
+	r1->nlimbs = k+1;
+    mpi_mul( r2, r2, m );
+    if( r2->nlimbs > k+1 ) /* quick modulo operation */
+	r2->nlimbs = k+1;
+    mpi_sub( r, r1, r2 );
+
+    if( mpi_is_neg( r ) ) {
+	MPI tmp;
+
+	tmp = mpi_alloc( k + 2 );
+	mpi_set_ui( tmp, 1 );
+	mpi_lshift_limbs( tmp, k+1 );
+	mpi_add( r, r, tmp );
+	mpi_free(tmp);
+    }
+
+    /* 4. while r >= m do r = r - m */
+    while( mpi_cmp( r, m ) >= 0 )
+	mpi_sub( r, r, m );
+
+    return 0;
+}
+#endif /* USE_BARRETT */
+
diff --git a/mpi/mpi-mul.c b/mpi/mpi-mul.c
index f83824926..df8eb2586 100644
--- a/mpi/mpi-mul.c
+++ b/mpi/mpi-mul.c
@@ -189,7 +189,6 @@ mpi_mul( MPI w, MPI u, MPI v)
 }
 
 
-
 void
 mpi_mulm( MPI w, MPI u, MPI v, MPI m)
 {
-- 
cgit v1.2.3