summaryrefslogtreecommitdiff
path: root/crypto/bn/bn_asm.c
diff options
context:
space:
mode:
Diffstat (limited to 'crypto/bn/bn_asm.c')
-rw-r--r--crypto/bn/bn_asm.c352
1 files changed, 275 insertions, 77 deletions
diff --git a/crypto/bn/bn_asm.c b/crypto/bn/bn_asm.c
index be8aa3ffc5..c43c91cc09 100644
--- a/crypto/bn/bn_asm.c
+++ b/crypto/bn/bn_asm.c
@@ -75,6 +75,7 @@ BN_ULONG bn_mul_add_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w)
assert(num >= 0);
if (num <= 0) return(c1);
+#ifndef OPENSSL_SMALL_FOOTPRINT
while (num&~3)
{
mul_add(rp[0],ap[0],w,c1);
@@ -83,11 +84,11 @@ BN_ULONG bn_mul_add_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w)
mul_add(rp[3],ap[3],w,c1);
ap+=4; rp+=4; num-=4;
}
- if (num)
+#endif
+ while (num)
{
- mul_add(rp[0],ap[0],w,c1); if (--num==0) return c1;
- mul_add(rp[1],ap[1],w,c1); if (--num==0) return c1;
- mul_add(rp[2],ap[2],w,c1); return c1;
+ mul_add(rp[0],ap[0],w,c1);
+ ap++; rp++; num--;
}
return(c1);
@@ -100,6 +101,7 @@ BN_ULONG bn_mul_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w)
assert(num >= 0);
if (num <= 0) return(c1);
+#ifndef OPENSSL_SMALL_FOOTPRINT
while (num&~3)
{
mul(rp[0],ap[0],w,c1);
@@ -108,11 +110,11 @@ BN_ULONG bn_mul_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w)
mul(rp[3],ap[3],w,c1);
ap+=4; rp+=4; num-=4;
}
- if (num)
+#endif
+ while (num)
{
- mul(rp[0],ap[0],w,c1); if (--num == 0) return c1;
- mul(rp[1],ap[1],w,c1); if (--num == 0) return c1;
- mul(rp[2],ap[2],w,c1);
+ mul(rp[0],ap[0],w,c1);
+ ap++; rp++; num--;
}
return(c1);
}
@@ -121,6 +123,8 @@ void bn_sqr_words(BN_ULONG *r, const BN_ULONG *a, int n)
{
assert(n >= 0);
if (n <= 0) return;
+
+#ifndef OPENSSL_SMALL_FOOTPRINT
while (n&~3)
{
sqr(r[0],r[1],a[0]);
@@ -129,11 +133,11 @@ void bn_sqr_words(BN_ULONG *r, const BN_ULONG *a, int n)
sqr(r[6],r[7],a[3]);
a+=4; r+=8; n-=4;
}
- if (n)
+#endif
+ while (n)
{
- sqr(r[0],r[1],a[0]); if (--n == 0) return;
- sqr(r[2],r[3],a[1]); if (--n == 0) return;
- sqr(r[4],r[5],a[2]);
+ sqr(r[0],r[1],a[0]);
+ a++; r+=2; n--;
}
}
@@ -150,18 +154,20 @@ BN_ULONG bn_mul_add_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w)
bl=LBITS(w);
bh=HBITS(w);
- for (;;)
+#ifndef OPENSSL_SMALL_FOOTPRINT
+ while (num&~3)
{
mul_add(rp[0],ap[0],bl,bh,c);
- if (--num == 0) break;
mul_add(rp[1],ap[1],bl,bh,c);
- if (--num == 0) break;
mul_add(rp[2],ap[2],bl,bh,c);
- if (--num == 0) break;
mul_add(rp[3],ap[3],bl,bh,c);
- if (--num == 0) break;
- ap+=4;
- rp+=4;
+ ap+=4; rp+=4; num-=4;
+ }
+#endif
+ while (num)
+ {
+ mul_add(rp[0],ap[0],bl,bh,c);
+ ap++; rp++; num--;
}
return(c);
}
@@ -177,18 +183,20 @@ BN_ULONG bn_mul_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w)
bl=LBITS(w);
bh=HBITS(w);
- for (;;)
+#ifndef OPENSSL_SMALL_FOOTPRINT
+ while (num&~3)
{
mul(rp[0],ap[0],bl,bh,carry);
- if (--num == 0) break;
mul(rp[1],ap[1],bl,bh,carry);
- if (--num == 0) break;
mul(rp[2],ap[2],bl,bh,carry);
- if (--num == 0) break;
mul(rp[3],ap[3],bl,bh,carry);
- if (--num == 0) break;
- ap+=4;
- rp+=4;
+ ap+=4; rp+=4; num-=4;
+ }
+#endif
+ while (num)
+ {
+ mul(rp[0],ap[0],bl,bh,carry);
+ ap++; rp++; num--;
}
return(carry);
}
@@ -197,22 +205,21 @@ void bn_sqr_words(BN_ULONG *r, const BN_ULONG *a, int n)
{
assert(n >= 0);
if (n <= 0) return;
- for (;;)
+
+#ifndef OPENSSL_SMALL_FOOTPRINT
+ while (n&~3)
{
sqr64(r[0],r[1],a[0]);
- if (--n == 0) break;
-
sqr64(r[2],r[3],a[1]);
- if (--n == 0) break;
-
sqr64(r[4],r[5],a[2]);
- if (--n == 0) break;
-
sqr64(r[6],r[7],a[3]);
- if (--n == 0) break;
-
- a+=4;
- r+=8;
+ a+=4; r+=8; n-=4;
+ }
+#endif
+ while (n)
+ {
+ sqr64(r[0],r[1],a[0]);
+ a++; r+=2; n--;
}
}
@@ -237,7 +244,7 @@ BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d)
if (d == 0) return(BN_MASK2);
i=BN_num_bits_word(d);
- assert((i == BN_BITS2) || (h > (BN_ULONG)1<<i));
+ assert((i == BN_BITS2) || (h <= (BN_ULONG)1<<i));
i=BN_BITS2-i;
if (h >= d) h-=d;
@@ -303,31 +310,30 @@ BN_ULONG bn_add_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b, int n)
assert(n >= 0);
if (n <= 0) return((BN_ULONG)0);
- for (;;)
+#ifndef OPENSSL_SMALL_FOOTPRINT
+ while (n&~3)
{
ll+=(BN_ULLONG)a[0]+b[0];
r[0]=(BN_ULONG)ll&BN_MASK2;
ll>>=BN_BITS2;
- if (--n <= 0) break;
-
ll+=(BN_ULLONG)a[1]+b[1];
r[1]=(BN_ULONG)ll&BN_MASK2;
ll>>=BN_BITS2;
- if (--n <= 0) break;
-
ll+=(BN_ULLONG)a[2]+b[2];
r[2]=(BN_ULONG)ll&BN_MASK2;
ll>>=BN_BITS2;
- if (--n <= 0) break;
-
ll+=(BN_ULLONG)a[3]+b[3];
r[3]=(BN_ULONG)ll&BN_MASK2;
ll>>=BN_BITS2;
- if (--n <= 0) break;
-
- a+=4;
- b+=4;
- r+=4;
+ a+=4; b+=4; r+=4; n-=4;
+ }
+#endif
+ while (n)
+ {
+ ll+=(BN_ULLONG)a[0]+b[0];
+ r[0]=(BN_ULONG)ll&BN_MASK2;
+ ll>>=BN_BITS2;
+ a++; b++; r++; n--;
}
return((BN_ULONG)ll);
}
@@ -340,7 +346,8 @@ BN_ULONG bn_add_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b, int n)
if (n <= 0) return((BN_ULONG)0);
c=0;
- for (;;)
+#ifndef OPENSSL_SMALL_FOOTPRINT
+ while (n&~3)
{
t=a[0];
t=(t+c)&BN_MASK2;
@@ -348,35 +355,36 @@ BN_ULONG bn_add_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b, int n)
l=(t+b[0])&BN_MASK2;
c+=(l < t);
r[0]=l;
- if (--n <= 0) break;
-
t=a[1];
t=(t+c)&BN_MASK2;
c=(t < c);
l=(t+b[1])&BN_MASK2;
c+=(l < t);
r[1]=l;
- if (--n <= 0) break;
-
t=a[2];
t=(t+c)&BN_MASK2;
c=(t < c);
l=(t+b[2])&BN_MASK2;
c+=(l < t);
r[2]=l;
- if (--n <= 0) break;
-
t=a[3];
t=(t+c)&BN_MASK2;
c=(t < c);
l=(t+b[3])&BN_MASK2;
c+=(l < t);
r[3]=l;
- if (--n <= 0) break;
-
- a+=4;
- b+=4;
- r+=4;
+ a+=4; b+=4; r+=4; n-=4;
+ }
+#endif
+ while(n)
+ {
+ t=a[0];
+ t=(t+c)&BN_MASK2;
+ c=(t < c);
+ l=(t+b[0])&BN_MASK2;
+ c+=(l < t);
+ r[0]=l;
+ a++; b++; r++; n--;
}
return((BN_ULONG)c);
}
@@ -390,36 +398,35 @@ BN_ULONG bn_sub_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b, int n)
assert(n >= 0);
if (n <= 0) return((BN_ULONG)0);
- for (;;)
+#ifndef OPENSSL_SMALL_FOOTPRINT
+ while (n&~3)
{
t1=a[0]; t2=b[0];
r[0]=(t1-t2-c)&BN_MASK2;
if (t1 != t2) c=(t1 < t2);
- if (--n <= 0) break;
-
t1=a[1]; t2=b[1];
r[1]=(t1-t2-c)&BN_MASK2;
if (t1 != t2) c=(t1 < t2);
- if (--n <= 0) break;
-
t1=a[2]; t2=b[2];
r[2]=(t1-t2-c)&BN_MASK2;
if (t1 != t2) c=(t1 < t2);
- if (--n <= 0) break;
-
t1=a[3]; t2=b[3];
r[3]=(t1-t2-c)&BN_MASK2;
if (t1 != t2) c=(t1 < t2);
- if (--n <= 0) break;
-
- a+=4;
- b+=4;
- r+=4;
+ a+=4; b+=4; r+=4; n-=4;
+ }
+#endif
+ while (n)
+ {
+ t1=a[0]; t2=b[0];
+ r[0]=(t1-t2-c)&BN_MASK2;
+ if (t1 != t2) c=(t1 < t2);
+ a++; b++; r++; n--;
}
return(c);
}
-#ifdef BN_MUL_COMBA
+#if defined(BN_MUL_COMBA) && !defined(OPENSSL_SMALL_FOOTPRINT)
#undef bn_mul_comba8
#undef bn_mul_comba4
@@ -459,6 +466,34 @@ BN_ULONG bn_sub_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b, int n)
#define sqr_add_c2(a,i,j,c0,c1,c2) \
mul_add_c2((a)[i],(a)[j],c0,c1,c2)
+#elif defined(BN_UMULT_LOHI)
+
+#define mul_add_c(a,b,c0,c1,c2) { \
+ BN_ULONG ta=(a),tb=(b); \
+ BN_UMULT_LOHI(t1,t2,ta,tb); \
+ c0 += t1; t2 += (c0<t1)?1:0; \
+ c1 += t2; c2 += (c1<t2)?1:0; \
+ }
+
+#define mul_add_c2(a,b,c0,c1,c2) { \
+ BN_ULONG ta=(a),tb=(b),t0; \
+ BN_UMULT_LOHI(t0,t1,ta,tb); \
+ t2 = t1+t1; c2 += (t2<t1)?1:0; \
+ t1 = t0+t0; t2 += (t1<t0)?1:0; \
+ c0 += t1; t2 += (c0<t1)?1:0; \
+ c1 += t2; c2 += (c1<t2)?1:0; \
+ }
+
+#define sqr_add_c(a,i,c0,c1,c2) { \
+ BN_ULONG ta=(a)[i]; \
+ BN_UMULT_LOHI(t1,t2,ta,ta); \
+ c0 += t1; t2 += (c0<t1)?1:0; \
+ c1 += t2; c2 += (c1<t2)?1:0; \
+ }
+
+#define sqr_add_c2(a,i,j,c0,c1,c2) \
+ mul_add_c2((a)[i],(a)[j],c0,c1,c2)
+
#elif defined(BN_UMULT_HIGH)
#define mul_add_c(a,b,c0,c1,c2) { \
@@ -792,18 +827,134 @@ void bn_sqr_comba4(BN_ULONG *r, const BN_ULONG *a)
r[6]=c1;
r[7]=c2;
}
+
+#ifdef OPENSSL_NO_ASM
+#ifdef OPENSSL_BN_ASM_MONT
+#include <alloca.h>
+/*
+ * This is essentially reference implementation, which may or may not
+ * result in performance improvement. E.g. on IA-32 this routine was
+ * observed to give 40% faster rsa1024 private key operations and 10%
+ * faster rsa4096 ones, while on AMD64 it improves rsa1024 sign only
+ * by 10% and *worsens* rsa4096 sign by 15%. Once again, it's a
+ * reference implementation, one to be used as starting point for
+ * platform-specific assembler. Mentioned numbers apply to compiler
+ * generated code compiled with and without -DOPENSSL_BN_ASM_MONT and
+ * can vary not only from platform to platform, but even for compiler
+ * versions. Assembler vs. assembler improvement coefficients can
+ * [and are known to] differ and are to be documented elsewhere.
+ */
+int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, const BN_ULONG *np,const BN_ULONG *n0p, int num)
+ {
+ BN_ULONG c0,c1,ml,*tp,n0;
+#ifdef mul64
+ BN_ULONG mh;
+#endif
+ volatile BN_ULONG *vp;
+ int i=0,j;
+
+#if 0 /* template for platform-specific implementation */
+ if (ap==bp) return bn_sqr_mont(rp,ap,np,n0p,num);
+#endif
+ vp = tp = alloca((num+2)*sizeof(BN_ULONG));
+
+ n0 = *n0p;
+
+ c0 = 0;
+ ml = bp[0];
+#ifdef mul64
+ mh = HBITS(ml);
+ ml = LBITS(ml);
+ for (j=0;j<num;++j)
+ mul(tp[j],ap[j],ml,mh,c0);
+#else
+ for (j=0;j<num;++j)
+ mul(tp[j],ap[j],ml,c0);
+#endif
+
+ tp[num] = c0;
+ tp[num+1] = 0;
+ goto enter;
+
+ for(i=0;i<num;i++)
+ {
+ c0 = 0;
+ ml = bp[i];
+#ifdef mul64
+ mh = HBITS(ml);
+ ml = LBITS(ml);
+ for (j=0;j<num;++j)
+ mul_add(tp[j],ap[j],ml,mh,c0);
+#else
+ for (j=0;j<num;++j)
+ mul_add(tp[j],ap[j],ml,c0);
+#endif
+ c1 = (tp[num] + c0)&BN_MASK2;
+ tp[num] = c1;
+ tp[num+1] = (c1<c0?1:0);
+ enter:
+ c1 = tp[0];
+ ml = (c1*n0)&BN_MASK2;
+ c0 = 0;
+#ifdef mul64
+ mh = HBITS(ml);
+ ml = LBITS(ml);
+ mul_add(c1,np[0],ml,mh,c0);
+#else
+ mul_add(c1,ml,np[0],c0);
+#endif
+ for(j=1;j<num;j++)
+ {
+ c1 = tp[j];
+#ifdef mul64
+ mul_add(c1,np[j],ml,mh,c0);
+#else
+ mul_add(c1,ml,np[j],c0);
+#endif
+ tp[j-1] = c1&BN_MASK2;
+ }
+ c1 = (tp[num] + c0)&BN_MASK2;
+ tp[num-1] = c1;
+ tp[num] = tp[num+1] + (c1<c0?1:0);
+ }
+
+ if (tp[num]!=0 || tp[num-1]>=np[num-1])
+ {
+ c0 = bn_sub_words(rp,tp,np,num);
+ if (tp[num]!=0 || c0==0)
+ {
+ for(i=0;i<num+2;i++) vp[i] = 0;
+ return 1;
+ }
+ }
+ for(i=0;i<num;i++) rp[i] = tp[i], vp[i] = 0;
+ vp[num] = 0;
+ vp[num+1] = 0;
+ return 1;
+ }
+#else
+/*
+ * Return value of 0 indicates that multiplication/convolution was not
+ * performed to signal the caller to fall down to alternative/original
+ * code-path.
+ */
+int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, const BN_ULONG *np,const BN_ULONG *n0, int num)
+{ return 0; }
+#endif /* OPENSSL_BN_ASM_MONT */
+#endif
+
#else /* !BN_MUL_COMBA */
/* hmm... is it faster just to do a multiply? */
#undef bn_sqr_comba4
-void bn_sqr_comba4(BN_ULONG *r, BN_ULONG *a)
+void bn_sqr_comba4(BN_ULONG *r, const BN_ULONG *a)
{
BN_ULONG t[8];
bn_sqr_normal(r,a,4,t);
}
#undef bn_sqr_comba8
-void bn_sqr_comba8(BN_ULONG *r, BN_ULONG *a)
+void bn_sqr_comba8(BN_ULONG *r, const BN_ULONG *a)
{
BN_ULONG t[16];
bn_sqr_normal(r,a,8,t);
@@ -829,4 +980,51 @@ void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
r[15]=bn_mul_add_words(&(r[7]),a,8,b[7]);
}
+#ifdef OPENSSL_NO_ASM
+#ifdef OPENSSL_BN_ASM_MONT
+#include <alloca.h>
+int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, const BN_ULONG *np,const BN_ULONG *n0p, int num)
+ {
+ BN_ULONG c0,c1,*tp,n0=*n0p;
+ volatile BN_ULONG *vp;
+ int i=0,j;
+
+ vp = tp = alloca((num+2)*sizeof(BN_ULONG));
+
+ for(i=0;i<=num;i++) tp[i]=0;
+
+ for(i=0;i<num;i++)
+ {
+ c0 = bn_mul_add_words(tp,ap,num,bp[i]);
+ c1 = (tp[num] + c0)&BN_MASK2;
+ tp[num] = c1;
+ tp[num+1] = (c1<c0?1:0);
+
+ c0 = bn_mul_add_words(tp,np,num,tp[0]*n0);
+ c1 = (tp[num] + c0)&BN_MASK2;
+ tp[num] = c1;
+ tp[num+1] += (c1<c0?1:0);
+ for(j=0;j<=num;j++) tp[j]=tp[j+1];
+ }
+
+ if (tp[num]!=0 || tp[num-1]>=np[num-1])
+ {
+ c0 = bn_sub_words(rp,tp,np,num);
+ if (tp[num]!=0 || c0==0)
+ {
+ for(i=0;i<num+2;i++) vp[i] = 0;
+ return 1;
+ }
+ }
+ for(i=0;i<num;i++) rp[i] = tp[i], vp[i] = 0;
+ vp[num] = 0;
+ vp[num+1] = 0;
+ return 1;
+ }
+#else
+int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, const BN_ULONG *np,const BN_ULONG *n0, int num)
+{ return 0; }
+#endif /* OPENSSL_BN_ASM_MONT */
+#endif
+
#endif /* !BN_MUL_COMBA */
View Lorry Controller Status