X-Git-Url: https://git.openssl.org/?p=openssl.git;a=blobdiff_plain;f=crypto%2Fbn%2Fbn_exp.c;h=070fd31f92eecd706647d4383b42bf9092f34db0;hp=8f8c694481911cdc8a71115bdc0ef65abfa555bc;hb=0600a5cd4996e8e469b99470c8eda24498ab4373;hpb=46a643763de6d8e39ecf6f76fa79b4d04885aa59

diff --git a/crypto/bn/bn_exp.c b/crypto/bn/bn_exp.c
index 8f8c694481..070fd31f92 100644
--- a/crypto/bn/bn_exp.c
+++ b/crypto/bn/bn_exp.c
@@ -113,6 +113,35 @@
 #include "cryptlib.h"
 #include "bn_lcl.h"
 
+#include <stdlib.h>
+#ifdef _WIN32
+# include <malloc.h>
+# ifndef alloca
+#  define alloca _alloca
+# endif
+#elif defined(__GNUC__)
+# ifndef alloca
+#  define alloca(s) __builtin_alloca((s))
+# endif
+#elif defined(__sun)
+# include <alloca.h>
+#endif
+
+#undef RSAZ_ENABLED
+#if defined(OPENSSL_BN_ASM_MONT) && \
+	(defined(__x86_64) || defined(__x86_64__) || \
+	 defined(_M_AMD64) || defined(_M_X64))
+# include "rsaz_exp.h"
+# define RSAZ_ENABLED
+#endif
+
+#undef SPARC_T4_MONT
+#if defined(OPENSSL_BN_ASM_MONT) && (defined(__sparc__) || defined(__sparc))
+# include "sparc_arch.h"
+extern unsigned int OPENSSL_sparcv9cap_P[];
+# define SPARC_T4_MONT
+#endif
+
 /* maximum precomputation table size for *variable* sliding windows */
 #define TABLE_SIZE	32
 
@@ -122,9 +151,9 @@ int BN_exp(BIGNUM *r, const BIGNUM *a, const BIGNUM *p, BN_CTX *ctx)
 	int i,bits,ret=0;
 	BIGNUM *v,*rr;
 
-	if (BN_get_flags(p, BN_FLG_EXP_CONSTTIME) != 0)
+	if (BN_get_flags(p, BN_FLG_CONSTTIME) != 0)
 		{
-		/* BN_FLG_EXP_CONSTTIME only supported by BN_mod_exp_mont() */
+		/* BN_FLG_CONSTTIME only supported by BN_mod_exp_mont() */
 		BNerr(BN_F_BN_EXP,ERR_R_SHOULD_NOT_HAVE_BEEN_CALLED);
 		return -1;
 		}
@@ -134,7 +163,8 @@ int BN_exp(BIGNUM *r, const BIGNUM *a, const BIGNUM *p, BN_CTX *ctx)
 		rr = BN_CTX_get(ctx);
 	else
 		rr = r;
-	if ((v = BN_CTX_get(ctx)) == NULL) goto err;
+	v = BN_CTX_get(ctx);
+	if (rr == NULL || v == NULL) goto err;
 
 	if (BN_copy(v,a) == NULL) goto err;
 	bits=BN_num_bits(p);
@@ -213,7 +243,7 @@ int BN_mod_exp(BIGNUM *r, const BIGNUM *a, const BIGNUM *p, const BIGNUM *m,
 	if (BN_is_odd(m))
 		{
 #  ifdef MONT_EXP_WORD
-		if (a->top == 1 && !a->neg && (BN_get_flags(p, BN_FLG_EXP_CONSTTIME) == 0))
+		if (a->top == 1 && !a->neg && (BN_get_flags(p, BN_FLG_CONSTTIME) == 0))
 			{
 			BN_ULONG A = a->d[0];
 			ret=BN_mod_exp_mont_word(r,A,p,m,ctx,NULL);
@@ -245,9 +275,9 @@ int BN_mod_exp_recp(BIGNUM *r, const BIGNUM *a, const BIGNUM *p,
 	BIGNUM *val[TABLE_SIZE];
 	BN_RECP_CTX recp;
 
-	if (BN_get_flags(p, BN_FLG_EXP_CONSTTIME) != 0)
+	if (BN_get_flags(p, BN_FLG_CONSTTIME) != 0)
 		{
-		/* BN_FLG_EXP_CONSTTIME only supported by BN_mod_exp_mont() */
+		/* BN_FLG_CONSTTIME only supported by BN_mod_exp_mont() */
 		BNerr(BN_F_BN_MOD_EXP_RECP,ERR_R_SHOULD_NOT_HAVE_BEEN_CALLED);
 		return -1;
 		}
@@ -379,7 +409,7 @@ int BN_mod_exp_mont(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p,
 	BIGNUM *val[TABLE_SIZE];
 	BN_MONT_CTX *mont=NULL;
 
-	if (BN_get_flags(p, BN_FLG_EXP_CONSTTIME) != 0)
+	if (BN_get_flags(p, BN_FLG_CONSTTIME) != 0)
 		{
 		return BN_mod_exp_mont_consttime(rr, a, p, m, ctx, in_mont);
 		}
@@ -454,6 +484,21 @@ int BN_mod_exp_mont(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p,
 	wstart=bits-1;	/* The top bit of the window */
 	wend=0;		/* The bottom bit of the window */
 
+#if 1	/* by Shay Gueron's suggestion */
+	j = m->top;	/* borrow j */
+	if (m->d[j-1] & (((BN_ULONG)1)<<(BN_BITS2-1)))
+		{
+		if (bn_wexpand(r,j) == NULL) goto err;
+		/* 2^(top*BN_BITS2) - m */
+		r->d[0] = (0-m->d[0])&BN_MASK2;
+		for(i=1;i<j;i++) r->d[i] = (~m->d[i])&BN_MASK2;
+		r->top = j;
+		/* Upper words will be zero if the corresponding words of 'm'
+		 * were 0xfff[...], so decrement r->top accordingly. */
+		bn_correct_top(r);
+		}
+	else
+#endif
 	if (!BN_to_montgomery(r,BN_value_one(),mont,ctx)) goto err;
 	for (;;)
 		{
@@ -506,6 +551,17 @@ int BN_mod_exp_mont(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p,
 		start=0;
 		if (wstart < 0) break;
 		}
+#if defined(SPARC_T4_MONT)
+	if (OPENSSL_sparcv9cap_P[0]&(SPARCV9_VIS3|SPARCV9_PREFER_FPU))
+		{
+		j = mont->N.top;	/* borrow j */
+		val[0]->d[0] = 1;	/* borrow val[0] */
+		for (i=1;i<j;i++) val[0]->d[i] = 0;
+		val[0]->top = j;
+		if (!BN_mod_mul_montgomery(rr,r,val[0],mont,ctx)) goto err;
+		}
+	else
+#endif
 	if (!BN_from_montgomery(rr,r,mont,ctx)) goto err;
 	ret=1;
 err:
@@ -515,29 +571,45 @@ err:
 	return(ret);
 	}
 
+#if defined(SPARC_T4_MONT)
+static BN_ULONG bn_get_bits(const BIGNUM *a, int bitpos)
+	{
+	BN_ULONG ret=0;
+	int wordpos;
+
+	wordpos = bitpos/BN_BITS2;
+	bitpos %= BN_BITS2;
+	if (wordpos>=0 && wordpos < a->top)
+		{
+		ret = a->d[wordpos]&BN_MASK2;
+		if (bitpos)
+			{
+			ret >>= bitpos;
+			if (++wordpos < a->top)
+				ret |= a->d[wordpos]<<(BN_BITS2-bitpos);
+			}
+		}
+
+	return ret&BN_MASK2;
+}
+#endif
 
 /* BN_mod_exp_mont_consttime() stores the precomputed powers in a specific layout
  * so that accessing any of these table values shows the same access pattern as far
  * as cache lines are concerned.  The following functions are used to transfer a BIGNUM
  * from/to that table. */
 
-static int MOD_EXP_CTIME_COPY_TO_PREBUF(BIGNUM *b, int top, unsigned char *buf, int idx, int width)
+static int MOD_EXP_CTIME_COPY_TO_PREBUF(const BIGNUM *b, int top, unsigned char *buf, int idx, int width)
 	{
 	size_t i, j;
 
-	if (bn_wexpand(b, top) == NULL)
-		return 0;
-	while (b->top < top)
-		{
-		b->d[b->top++] = 0;
-		}
-	
+	if (top > b->top)
+		top = b->top; /* this works because 'buf' is explicitly zeroed */
 	for (i = 0, j=idx; i < top * sizeof b->d[0]; i++, j+=width)
 		{
 		buf[j] = ((unsigned char*)b->d)[i];
 		}
 
-	bn_correct_top(b);
 	return 1;
 	}
 
@@ -560,7 +632,7 @@ static int MOD_EXP_CTIME_COPY_FROM_PREBUF(BIGNUM *b, int top, unsigned char *buf
 
 /* Given a pointer value, compute the next address that is a cache line multiple. */
 #define MOD_EXP_CTIME_ALIGN(x_) \
-	((unsigned char*)(x_) + (MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH - (((BN_ULONG)(x_)) & (MOD_EXP_CTIME_MIN_CACHE_LINE_MASK))))
+	((unsigned char*)(x_) + (MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH - (((size_t)(x_)) & (MOD_EXP_CTIME_MIN_CACHE_LINE_MASK))))
 
 /* This variant of BN_mod_exp_mont() uses fixed windows and the special
  * precomputation memory layout to limit data-dependency to a minimum
@@ -571,17 +643,18 @@ static int MOD_EXP_CTIME_COPY_FROM_PREBUF(BIGNUM *b, int top, unsigned char *buf
 int BN_mod_exp_mont_consttime(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p,
 		    const BIGNUM *m, BN_CTX *ctx, BN_MONT_CTX *in_mont)
 	{
-	int i,bits,ret=0,idx,window,wvalue;
+	int i,bits,ret=0,window,wvalue;
 	int top;
- 	BIGNUM *r;
-	const BIGNUM *aa;
 	BN_MONT_CTX *mont=NULL;
 
 	int numPowers;
 	unsigned char *powerbufFree=NULL;
 	int powerbufLen = 0;
 	unsigned char *powerbuf=NULL;
-	BIGNUM *computeTemp=NULL, *am=NULL;
+	BIGNUM tmp, am;
+#if defined(SPARC_T4_MONT)
+	unsigned int t4=0;
+#endif
 
 	bn_check_top(a);
 	bn_check_top(p);
@@ -601,10 +674,7 @@ int BN_mod_exp_mont_consttime(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p,
 		return ret;
 		}
 
- 	/* Initialize BIGNUM context and allocate intermediate result */
 	BN_CTX_start(ctx);
-	r = BN_CTX_get(ctx);
-	if (r == NULL) goto err;
 
 	/* Allocate a montgomery context if it was not supplied by the caller.
 	 * If this is not done, things will break in the montgomery part.
@@ -617,42 +687,344 @@ int BN_mod_exp_mont_consttime(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p,
 		if (!BN_MONT_CTX_set(mont,m,ctx)) goto err;
 		}
 
+#ifdef RSAZ_ENABLED
+	/*
+	 * If the size of the operands allow it, perform the optimized
+	 * RSAZ exponentiation. For further information see
+	 * crypto/bn/rsaz_exp.c and accompanying assembly modules.
+	 */
+	if ((16 == a->top) && (16 == p->top) && (BN_num_bits(m) == 1024)
+	    && rsaz_avx2_eligible())
+	    	{
+		if (NULL == bn_wexpand(rr, 16)) goto err;
+		RSAZ_1024_mod_exp_avx2(rr->d, a->d, p->d, m->d, mont->RR.d, mont->n0[0]);
+		rr->top = 16;
+		rr->neg = 0;
+		bn_correct_top(rr);
+		ret = 1;
+		goto err;
+		}
+	else if ((8 == a->top) && (8 == p->top) && (BN_num_bits(m) == 512))
+		{
+		if (NULL == bn_wexpand(rr,8)) goto err;
+		RSAZ_512_mod_exp(rr->d, a->d, p->d, m->d, mont->n0[0], mont->RR.d);
+		rr->top = 8;
+		rr->neg = 0;
+		bn_correct_top(rr);
+		ret = 1;
+		goto err;
+		}
+#endif
+
 	/* Get the window size to use with size of p. */
 	window = BN_window_bits_for_ctime_exponent_size(bits);
+#if defined(SPARC_T4_MONT)
+	if (window>=5 && (top&15)==0 && top<=64 &&
+	    (OPENSSL_sparcv9cap_P[1]&(CFR_MONTMUL|CFR_MONTSQR))==
+	    			     (CFR_MONTMUL|CFR_MONTSQR) &&
+	    (t4=OPENSSL_sparcv9cap_P[0]))
+		window=5;
+	else
+#endif
+#if defined(OPENSSL_BN_ASM_MONT5)
+	if (window>=5)
+		{
+		window=5;	/* ~5% improvement for RSA2048 sign, and even for RSA4096 */
+		if ((top&7)==0)	powerbufLen += 2*top*sizeof(m->d[0]);
+		}
+#endif
+	(void)0;
 
 	/* Allocate a buffer large enough to hold all of the pre-computed
-	 * powers of a.
+	 * powers of am, am itself and tmp.
 	 */
 	numPowers = 1 << window;
-	powerbufLen = sizeof(m->d[0])*top*numPowers;
+	powerbufLen += sizeof(m->d[0])*(top*numPowers +
+				((2*top)>numPowers?(2*top):numPowers));
+#ifdef alloca
+	if (powerbufLen < 3072)
+		powerbufFree = alloca(powerbufLen+MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH);
+	else
+#endif
 	if ((powerbufFree=(unsigned char*)OPENSSL_malloc(powerbufLen+MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH)) == NULL)
 		goto err;
 		
 	powerbuf = MOD_EXP_CTIME_ALIGN(powerbufFree);
 	memset(powerbuf, 0, powerbufLen);
 
- 	/* Initialize the intermediate result. Do this early to save double conversion,
-	 * once each for a^0 and intermediate result.
-	 */
- 	if (!BN_to_montgomery(r,BN_value_one(),mont,ctx)) goto err;
-	if (!MOD_EXP_CTIME_COPY_TO_PREBUF(r, top, powerbuf, 0, numPowers)) goto err;
+#ifdef alloca
+	if (powerbufLen < 3072)
+		powerbufFree = NULL;
+#endif
 
-	/* Initialize computeTemp as a^1 with montgomery precalcs */
-	computeTemp = BN_CTX_get(ctx);
-	am = BN_CTX_get(ctx);
-	if (computeTemp==NULL || am==NULL) goto err;
+	/* lay down tmp and am right after powers table */
+	tmp.d     = (BN_ULONG *)(powerbuf + sizeof(m->d[0])*top*numPowers);
+	am.d      = tmp.d + top;
+	tmp.top   = am.top  = 0;
+	tmp.dmax  = am.dmax = top;
+	tmp.neg   = am.neg  = 0;
+	tmp.flags = am.flags = BN_FLG_STATIC_DATA;
+
+	/* prepare a^0 in Montgomery domain */
+#if 1	/* by Shay Gueron's suggestion */
+	if (m->d[top-1] & (((BN_ULONG)1)<<(BN_BITS2-1)))
+		{
+		/* 2^(top*BN_BITS2) - m */
+		tmp.d[0] = (0-m->d[0])&BN_MASK2;
+		for (i=1;i<top;i++) tmp.d[i] = (~m->d[i])&BN_MASK2;
+		tmp.top = top;
+		}
+	else
+#endif
+ 	if (!BN_to_montgomery(&tmp,BN_value_one(),mont,ctx))	goto err;
 
+	/* prepare a^1 in Montgomery domain */
 	if (a->neg || BN_ucmp(a,m) >= 0)
 		{
-		if (!BN_mod(am,a,m,ctx))
-			goto err;
-		aa= am;
+		if (!BN_mod(&am,a,m,ctx))			goto err;
+		if (!BN_to_montgomery(&am,&am,mont,ctx))	goto err;
 		}
+	else	if (!BN_to_montgomery(&am,a,mont,ctx))		goto err;
+
+#if defined(SPARC_T4_MONT)
+    if (t4)
+	{
+	typedef int (*bn_pwr5_mont_f)(BN_ULONG *tp,const BN_ULONG *np,
+			const BN_ULONG *n0,const void *table,int power,int bits);
+	int bn_pwr5_mont_t4_8(BN_ULONG *tp,const BN_ULONG *np,
+			const BN_ULONG *n0,const void *table,int power,int bits);
+	int bn_pwr5_mont_t4_16(BN_ULONG *tp,const BN_ULONG *np,
+			const BN_ULONG *n0,const void *table,int power,int bits);
+	int bn_pwr5_mont_t4_24(BN_ULONG *tp,const BN_ULONG *np,
+			const BN_ULONG *n0,const void *table,int power,int bits);
+	int bn_pwr5_mont_t4_32(BN_ULONG *tp,const BN_ULONG *np,
+			const BN_ULONG *n0,const void *table,int power,int bits);
+	static const bn_pwr5_mont_f pwr5_funcs[4] = {
+			bn_pwr5_mont_t4_8,	bn_pwr5_mont_t4_16,
+			bn_pwr5_mont_t4_24,	bn_pwr5_mont_t4_32 };
+	bn_pwr5_mont_f pwr5_worker = pwr5_funcs[top/16-1];
+
+	typedef int (*bn_mul_mont_f)(BN_ULONG *rp,const BN_ULONG *ap,
+			const void *bp,const BN_ULONG *np,const BN_ULONG *n0);
+	int bn_mul_mont_t4_8(BN_ULONG *rp,const BN_ULONG *ap,
+			const void *bp,const BN_ULONG *np,const BN_ULONG *n0);
+	int bn_mul_mont_t4_16(BN_ULONG *rp,const BN_ULONG *ap,
+			const void *bp,const BN_ULONG *np,const BN_ULONG *n0);
+	int bn_mul_mont_t4_24(BN_ULONG *rp,const BN_ULONG *ap,
+			const void *bp,const BN_ULONG *np,const BN_ULONG *n0);
+	int bn_mul_mont_t4_32(BN_ULONG *rp,const BN_ULONG *ap,
+			const void *bp,const BN_ULONG *np,const BN_ULONG *n0);
+	static const bn_mul_mont_f mul_funcs[4] = {
+			bn_mul_mont_t4_8,	bn_mul_mont_t4_16,
+			bn_mul_mont_t4_24,	bn_mul_mont_t4_32 };
+	bn_mul_mont_f mul_worker = mul_funcs[top/16-1];
+
+	void bn_mul_mont_vis3(BN_ULONG *rp,const BN_ULONG *ap,
+			const void *bp,const BN_ULONG *np,
+			const BN_ULONG *n0,int num);
+	void bn_mul_mont_t4(BN_ULONG *rp,const BN_ULONG *ap,
+			const void *bp,const BN_ULONG *np,
+			const BN_ULONG *n0,int num);
+	void bn_mul_mont_gather5_t4(BN_ULONG *rp,const BN_ULONG *ap,
+			const void *table,const BN_ULONG *np,
+			const BN_ULONG *n0,int num,int power);
+	void bn_flip_n_scatter5_t4(const BN_ULONG *inp,size_t num,
+			void *table,size_t power);
+	void bn_gather5_t4(BN_ULONG *out,size_t num,
+			void *table,size_t power);
+	void bn_flip_t4(BN_ULONG *dst,BN_ULONG *src,size_t num);
+
+	BN_ULONG *np=mont->N.d, *n0=mont->n0;
+	int stride = 5*(6-(top/16-1));	/* multiple of 5, but less than 32 */
+
+	/* BN_to_montgomery can contaminate words above .top
+	 * [in BN_DEBUG[_DEBUG] build]... */
+	for (i=am.top; i<top; i++)	am.d[i]=0;
+	for (i=tmp.top; i<top; i++)	tmp.d[i]=0;
+
+	bn_flip_n_scatter5_t4(tmp.d,top,powerbuf,0);
+	bn_flip_n_scatter5_t4(am.d,top,powerbuf,1);
+	if (!(*mul_worker)(tmp.d,am.d,am.d,np,n0) &&
+	    !(*mul_worker)(tmp.d,am.d,am.d,np,n0))
+		bn_mul_mont_vis3(tmp.d,am.d,am.d,np,n0,top);
+	bn_flip_n_scatter5_t4(tmp.d,top,powerbuf,2);
+
+	for (i=3; i<32; i++)
+		{
+		/* Calculate a^i = a^(i-1) * a */
+		if (!(*mul_worker)(tmp.d,tmp.d,am.d,np,n0) &&
+		    !(*mul_worker)(tmp.d,tmp.d,am.d,np,n0))
+			bn_mul_mont_vis3(tmp.d,tmp.d,am.d,np,n0,top);
+		bn_flip_n_scatter5_t4(tmp.d,top,powerbuf,i);
+		}
+
+	/* switch to 64-bit domain */ 
+	np = alloca(top*sizeof(BN_ULONG));
+	top /= 2;
+	bn_flip_t4(np,mont->N.d,top);
+
+	bits--;
+	for (wvalue=0, i=bits%5; i>=0; i--,bits--)
+		wvalue = (wvalue<<1)+BN_is_bit_set(p,bits);
+	bn_gather5_t4(tmp.d,top,powerbuf,wvalue);
+
+	/* Scan the exponent one window at a time starting from the most
+	 * significant bits.
+	 */
+	while (bits >= 0)
+		{
+		if (bits < stride) stride = bits+1;
+		bits -= stride;
+		wvalue = bn_get_bits(p,bits+1);
+
+		if ((*pwr5_worker)(tmp.d,np,n0,powerbuf,wvalue,stride)) continue;
+		/* retry once and fall back */
+		if ((*pwr5_worker)(tmp.d,np,n0,powerbuf,wvalue,stride)) continue;
+
+		bits += stride-5;
+		wvalue >>= stride-5;
+		wvalue &= 31;
+		bn_mul_mont_t4(tmp.d,tmp.d,tmp.d,np,n0,top);
+		bn_mul_mont_t4(tmp.d,tmp.d,tmp.d,np,n0,top);
+		bn_mul_mont_t4(tmp.d,tmp.d,tmp.d,np,n0,top);
+		bn_mul_mont_t4(tmp.d,tmp.d,tmp.d,np,n0,top);
+		bn_mul_mont_t4(tmp.d,tmp.d,tmp.d,np,n0,top);
+		bn_mul_mont_gather5_t4(tmp.d,tmp.d,powerbuf,np,n0,top,wvalue);
+		}
+
+	bn_flip_t4(tmp.d,tmp.d,top);
+	top *= 2;
+	/* back to 32-bit domain */
+	tmp.top=top;
+	bn_correct_top(&tmp);
+	OPENSSL_cleanse(np,top*sizeof(BN_ULONG));
+	}
+    else
+#endif
+#if defined(OPENSSL_BN_ASM_MONT5)
+    /* This optimization uses ideas from http://eprint.iacr.org/2011/239,
+     * specifically optimization of cache-timing attack countermeasures
+     * and pre-computation optimization. */
+
+    /* Dedicated window==4 case improves 512-bit RSA sign by ~15%, but as
+     * 512-bit RSA is hardly relevant, we omit it to spare size... */ 
+    if (window==5 && top>1)
+	{
+	void bn_mul_mont_gather5(BN_ULONG *rp,const BN_ULONG *ap,
+			const void *table,const BN_ULONG *np,
+			const BN_ULONG *n0,int num,int power);
+	void bn_scatter5(const BN_ULONG *inp,size_t num,
+			void *table,size_t power);
+	void bn_gather5(BN_ULONG *out,size_t num,
+			void *table,size_t power);
+	void bn_power5(BN_ULONG *rp,const BN_ULONG *ap,
+			const void *table,const BN_ULONG *np,
+			const BN_ULONG *n0,int num,int power);
+	int bn_get_bits5(const BN_ULONG *ap,int off);
+	int bn_from_montgomery(BN_ULONG *rp,const BN_ULONG *ap,
+			const BN_ULONG *not_used,const BN_ULONG *np,
+			const BN_ULONG *n0,int num);
+
+	BN_ULONG *np=mont->N.d, *n0=mont->n0, *np2;
+
+	/* BN_to_montgomery can contaminate words above .top
+	 * [in BN_DEBUG[_DEBUG] build]... */
+	for (i=am.top; i<top; i++)	am.d[i]=0;
+	for (i=tmp.top; i<top; i++)	tmp.d[i]=0;
+
+	if (top&7)
+		np2 = np;
 	else
-		aa=a;
-	if (!BN_to_montgomery(am,aa,mont,ctx)) goto err;
-	if (!BN_copy(computeTemp, am)) goto err;
-	if (!MOD_EXP_CTIME_COPY_TO_PREBUF(am, top, powerbuf, 1, numPowers)) goto err;
+		for (np2=am.d+top,i=0; i<top; i++) np2[2*i]=np[i];
+
+	bn_scatter5(tmp.d,top,powerbuf,0);
+	bn_scatter5(am.d,am.top,powerbuf,1);
+	bn_mul_mont(tmp.d,am.d,am.d,np,n0,top);
+	bn_scatter5(tmp.d,top,powerbuf,2);
+
+#if 0
+	for (i=3; i<32; i++)
+		{
+		/* Calculate a^i = a^(i-1) * a */
+		bn_mul_mont_gather5(tmp.d,am.d,powerbuf,np2,n0,top,i-1);
+		bn_scatter5(tmp.d,top,powerbuf,i);
+		}
+#else
+	/* same as above, but uses squaring for 1/2 of operations */
+	for (i=4; i<32; i*=2)
+		{
+		bn_mul_mont(tmp.d,tmp.d,tmp.d,np,n0,top);
+		bn_scatter5(tmp.d,top,powerbuf,i);
+		}
+	for (i=3; i<8; i+=2)
+		{
+		int j;
+		bn_mul_mont_gather5(tmp.d,am.d,powerbuf,np2,n0,top,i-1);
+		bn_scatter5(tmp.d,top,powerbuf,i);
+		for (j=2*i; j<32; j*=2)
+			{
+			bn_mul_mont(tmp.d,tmp.d,tmp.d,np,n0,top);
+			bn_scatter5(tmp.d,top,powerbuf,j);
+			}
+		}
+	for (; i<16; i+=2)
+		{
+		bn_mul_mont_gather5(tmp.d,am.d,powerbuf,np2,n0,top,i-1);
+		bn_scatter5(tmp.d,top,powerbuf,i);
+		bn_mul_mont(tmp.d,tmp.d,tmp.d,np,n0,top);
+		bn_scatter5(tmp.d,top,powerbuf,2*i);
+		}
+	for (; i<32; i+=2)
+		{
+		bn_mul_mont_gather5(tmp.d,am.d,powerbuf,np2,n0,top,i-1);
+		bn_scatter5(tmp.d,top,powerbuf,i);
+		}
+#endif
+	bits--;
+	for (wvalue=0, i=bits%5; i>=0; i--,bits--)
+		wvalue = (wvalue<<1)+BN_is_bit_set(p,bits);
+	bn_gather5(tmp.d,top,powerbuf,wvalue);
+
+	/* Scan the exponent one window at a time starting from the most
+	 * significant bits.
+	 */
+	if (top&7)
+	    while (bits >= 0)
+		{
+		for (wvalue=0, i=0; i<5; i++,bits--)
+			wvalue = (wvalue<<1)+BN_is_bit_set(p,bits);
+
+		bn_mul_mont(tmp.d,tmp.d,tmp.d,np,n0,top);
+		bn_mul_mont(tmp.d,tmp.d,tmp.d,np,n0,top);
+		bn_mul_mont(tmp.d,tmp.d,tmp.d,np,n0,top);
+		bn_mul_mont(tmp.d,tmp.d,tmp.d,np,n0,top);
+		bn_mul_mont(tmp.d,tmp.d,tmp.d,np,n0,top);
+		bn_mul_mont_gather5(tmp.d,tmp.d,powerbuf,np,n0,top,wvalue);
+		}
+	else
+	    {
+	    while (bits >= 0)
+		{
+		wvalue = bn_get_bits5(p->d,bits-4);
+		bits-=5;
+		bn_power5(tmp.d,tmp.d,powerbuf,np2,n0,top,wvalue);
+		}
+	    }
+
+	ret=bn_from_montgomery(tmp.d,tmp.d,NULL,np2,n0,top);
+	tmp.top=top;
+	bn_correct_top(&tmp);
+	if (ret)
+		{
+		if (!BN_copy(rr,&tmp)) ret=0;
+		goto err; /* non-zero ret means it's not error */
+		}
+	}
+    else
+#endif
+	{
+	if (!MOD_EXP_CTIME_COPY_TO_PREBUF(&tmp, top, powerbuf, 0, numPowers)) goto err;
+	if (!MOD_EXP_CTIME_COPY_TO_PREBUF(&am,  top, powerbuf, 1, numPowers)) goto err;
 
 	/* If the window size is greater than 1, then calculate
 	 * val[i=2..2^winsize-1]. Powers are computed as a*a^(i-1)
@@ -661,62 +1033,63 @@ int BN_mod_exp_mont_consttime(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p,
 	 */
 	if (window > 1)
 		{
-		for (i=2; i<numPowers; i++)
+		if (!BN_mod_mul_montgomery(&tmp,&am,&am,mont,ctx))	goto err;
+		if (!MOD_EXP_CTIME_COPY_TO_PREBUF(&tmp, top, powerbuf, 2, numPowers)) goto err;
+		for (i=3; i<numPowers; i++)
 			{
 			/* Calculate a^i = a^(i-1) * a */
-			if (!BN_mod_mul_montgomery(computeTemp,am,computeTemp,mont,ctx))
+			if (!BN_mod_mul_montgomery(&tmp,&am,&tmp,mont,ctx))
 				goto err;
-			if (!MOD_EXP_CTIME_COPY_TO_PREBUF(computeTemp, top, powerbuf, i, numPowers)) goto err;
+			if (!MOD_EXP_CTIME_COPY_TO_PREBUF(&tmp, top, powerbuf, i, numPowers)) goto err;
 			}
 		}
 
- 	/* Adjust the number of bits up to a multiple of the window size.
- 	 * If the exponent length is not a multiple of the window size, then
- 	 * this pads the most significant bits with zeros to normalize the
- 	 * scanning loop to there's no special cases.
- 	 *
- 	 * * NOTE: Making the window size a power of two less than the native
-	 * * word size ensures that the padded bits won't go past the last
- 	 * * word in the internal BIGNUM structure. Going past the end will
- 	 * * still produce the correct result, but causes a different branch
- 	 * * to be taken in the BN_is_bit_set function.
- 	 */
- 	bits = ((bits+window-1)/window)*window;
- 	idx=bits-1;	/* The top bit of the window */
-
- 	/* Scan the exponent one window at a time starting from the most
- 	 * significant bits.
- 	 */
- 	while (idx >= 0)
+	bits--;
+	for (wvalue=0, i=bits%window; i>=0; i--,bits--)
+		wvalue = (wvalue<<1)+BN_is_bit_set(p,bits);
+	if (!MOD_EXP_CTIME_COPY_FROM_PREBUF(&tmp,top,powerbuf,wvalue,numPowers)) goto err;
+ 
+	/* Scan the exponent one window at a time starting from the most
+	 * significant bits.
+	 */
+ 	while (bits >= 0)
   		{
  		wvalue=0; /* The 'value' of the window */
  		
  		/* Scan the window, squaring the result as we go */
- 		for (i=0; i<window; i++,idx--)
+ 		for (i=0; i<window; i++,bits--)
  			{
-			if (!BN_mod_mul_montgomery(r,r,r,mont,ctx))	goto err;
-			wvalue = (wvalue<<1)+BN_is_bit_set(p,idx);
+			if (!BN_mod_mul_montgomery(&tmp,&tmp,&tmp,mont,ctx))	goto err;
+			wvalue = (wvalue<<1)+BN_is_bit_set(p,bits);
   			}
  		
 		/* Fetch the appropriate pre-computed value from the pre-buf */
-		if (!MOD_EXP_CTIME_COPY_FROM_PREBUF(computeTemp, top, powerbuf, wvalue, numPowers)) goto err;
+		if (!MOD_EXP_CTIME_COPY_FROM_PREBUF(&am, top, powerbuf, wvalue, numPowers)) goto err;
 
  		/* Multiply the result into the intermediate result */
- 		if (!BN_mod_mul_montgomery(r,r,computeTemp,mont,ctx)) goto err;
+ 		if (!BN_mod_mul_montgomery(&tmp,&tmp,&am,mont,ctx)) goto err;
   		}
+	}
 
  	/* Convert the final result from montgomery to standard format */
-	if (!BN_from_montgomery(rr,r,mont,ctx)) goto err;
+#if defined(SPARC_T4_MONT)
+	if (OPENSSL_sparcv9cap_P[0]&(SPARCV9_VIS3|SPARCV9_PREFER_FPU))
+		{
+		am.d[0] = 1;	/* borrow am */
+		for (i=1;i<top;i++) am.d[i] = 0;
+		if (!BN_mod_mul_montgomery(rr,&tmp,&am,mont,ctx)) goto err;
+		}
+	else
+#endif
+	if (!BN_from_montgomery(rr,&tmp,mont,ctx)) goto err;
 	ret=1;
 err:
 	if ((in_mont == NULL) && (mont != NULL)) BN_MONT_CTX_free(mont);
 	if (powerbuf!=NULL)
 		{
 		OPENSSL_cleanse(powerbuf,powerbufLen);
-		OPENSSL_free(powerbufFree);
+		if (powerbufFree) OPENSSL_free(powerbufFree);
 		}
- 	if (am!=NULL) BN_clear(am);
- 	if (computeTemp!=NULL) BN_clear(computeTemp);
 	BN_CTX_end(ctx);
 	return(ret);
 	}
@@ -745,9 +1118,9 @@ int BN_mod_exp_mont_word(BIGNUM *rr, BN_ULONG a, const BIGNUM *p,
 #define BN_TO_MONTGOMERY_WORD(r, w, mont) \
 		(BN_set_word(r, (w)) && BN_to_montgomery(r, r, (mont), ctx))
 
-	if (BN_get_flags(p, BN_FLG_EXP_CONSTTIME) != 0)
+	if (BN_get_flags(p, BN_FLG_CONSTTIME) != 0)
 		{
-		/* BN_FLG_EXP_CONSTTIME only supported by BN_mod_exp_mont() */
+		/* BN_FLG_CONSTTIME only supported by BN_mod_exp_mont() */
 		BNerr(BN_F_BN_MOD_EXP_MONT_WORD,ERR_R_SHOULD_NOT_HAVE_BEEN_CALLED);
 		return -1;
 		}
@@ -766,7 +1139,14 @@ int BN_mod_exp_mont_word(BIGNUM *rr, BN_ULONG a, const BIGNUM *p,
 	bits = BN_num_bits(p);
 	if (bits == 0)
 		{
-		ret = BN_one(rr);
+		/* x**0 mod 1 is still zero. */
+		if (BN_is_one(m))
+			{
+			ret = 1;
+			BN_zero(rr);
+			}
+		else
+			ret = BN_one(rr);
 		return ret;
 		}
 	if (a == 0)
@@ -881,9 +1261,9 @@ int BN_mod_exp_simple(BIGNUM *r, const BIGNUM *a, const BIGNUM *p,
 	/* Table of variables obtained from 'ctx' */
 	BIGNUM *val[TABLE_SIZE];
 
-	if (BN_get_flags(p, BN_FLG_EXP_CONSTTIME) != 0)
+	if (BN_get_flags(p, BN_FLG_CONSTTIME) != 0)
 		{
-		/* BN_FLG_EXP_CONSTTIME only supported by BN_mod_exp_mont() */
+		/* BN_FLG_CONSTTIME only supported by BN_mod_exp_mont() */
 		BNerr(BN_F_BN_MOD_EXP_SIMPLE,ERR_R_SHOULD_NOT_HAVE_BEEN_CALLED);
 		return -1;
 		}
@@ -987,4 +1367,3 @@ err:
 	bn_check_top(r);
 	return(ret);
 	}
-