From 8329e2e776176ce6ff9d7c48c7182943875065bd Mon Sep 17 00:00:00 2001 From: Andy Polyakov Date: Mon, 17 Oct 2011 17:41:49 +0000 Subject: [PATCH] bn_exp.c: further optimizations using more ideas from http://eprint.iacr.org/2011/239. --- crypto/bn/asm/x86_64-mont5.pl | 83 +++++++++++++-- crypto/bn/bn_exp.c | 186 +++++++++++++++------------------- 2 files changed, 156 insertions(+), 113 deletions(-) diff --git a/crypto/bn/asm/x86_64-mont5.pl b/crypto/bn/asm/x86_64-mont5.pl index d6e4c63df1..057cda28aa 100755 --- a/crypto/bn/asm/x86_64-mont5.pl +++ b/crypto/bn/asm/x86_64-mont5.pl @@ -607,7 +607,6 @@ $code.=<<___; add $A[1],$N[1] # np[j]*m1+ap[j]*bp[i]+tp[j] lea 4($j),$j # j+=2 adc \$0,%rdx - mov $N[1],(%rsp) # tp[j-1] mov %rdx,$N[0] jmp .Linner4x .align 16 @@ -626,7 +625,7 @@ $code.=<<___; adc \$0,%rdx add $A[0],$N[0] adc \$0,%rdx - mov $N[0],-24(%rsp,$j,8) # tp[j-1] + mov $N[1],-32(%rsp,$j,8) # tp[j-1] mov %rdx,$N[1] mulq $m0 # ap[j]*bp[i] @@ -643,7 +642,7 @@ $code.=<<___; adc \$0,%rdx add $A[1],$N[1] adc \$0,%rdx - mov $N[1],-16(%rsp,$j,8) # tp[j-1] + mov $N[0],-24(%rsp,$j,8) # tp[j-1] mov %rdx,$N[0] mulq $m0 # ap[j]*bp[i] @@ -660,7 +659,7 @@ $code.=<<___; adc \$0,%rdx add $A[0],$N[0] adc \$0,%rdx - mov $N[0],-8(%rsp,$j,8) # tp[j-1] + mov $N[1],-16(%rsp,$j,8) # tp[j-1] mov %rdx,$N[1] mulq $m0 # ap[j]*bp[i] @@ -678,7 +677,7 @@ $code.=<<___; adc \$0,%rdx add $A[1],$N[1] adc \$0,%rdx - mov $N[1],-32(%rsp,$j,8) # tp[j-1] + mov $N[0],-40(%rsp,$j,8) # tp[j-1] mov %rdx,$N[0] cmp $num,$j jl .Linner4x @@ -697,7 +696,7 @@ $code.=<<___; adc \$0,%rdx add $A[0],$N[0] adc \$0,%rdx - mov $N[0],-24(%rsp,$j,8) # tp[j-1] + mov $N[1],-32(%rsp,$j,8) # tp[j-1] mov %rdx,$N[1] mulq $m0 # ap[j]*bp[i] @@ -715,10 +714,11 @@ $code.=<<___; adc \$0,%rdx add $A[1],$N[1] adc \$0,%rdx - mov $N[1],-16(%rsp,$j,8) # tp[j-1] + mov $N[0],-24(%rsp,$j,8) # tp[j-1] mov %rdx,$N[0] movq %xmm0,$m0 # bp[i+1] + mov $N[1],-16(%rsp,$j,8) # tp[j-1] xor $N[1],$N[1] add $A[0],$N[0] @@ -831,6 +831,10 @@ ___ { my ($inp,$num,$tbl,$idx)=$win64?("%rcx","%rdx","%r8", "%r9") : # Win64 order ("%rdi","%rsi","%rdx","%rcx"); # Unix order +my $out=$inp; +my $STRIDE=2**5*8; +my $N=$STRIDE/4; + $code.=<<___; .globl bn_scatter5 .type bn_scatter5,\@abi-omnipotent @@ -849,6 +853,61 @@ bn_scatter5: .Lscatter_epilogue: ret .size bn_scatter5,.-bn_scatter5 + +.globl bn_gather5 +.type bn_gather5,\@abi-omnipotent +.align 16 +bn_gather5: +___ +$code.=<<___ if ($win64); +.LSEH_begin_bn_gather5: + # I can't trust assembler to use specific encoding:-( + .byte 0x48,0x83,0xec,0x28 #sub \$0x28,%rsp + .byte 0x0f,0x29,0x34,0x24 #movaps %xmm6,(%rsp) + .byte 0x0f,0x29,0x7c,0x24,0x10 #movdqa %xmm7,0x10(%rsp) +___ +$code.=<<___; + mov $idx,%r11 + shr \$`log($N/8)/log(2)`,$idx + and \$`$N/8-1`,%r11 + not $idx + lea .Lmagic_masks(%rip),%rax + and \$`2**5/($N/8)-1`,$idx # 5 is "window size" + lea 96($tbl,%r11,8),$tbl # pointer within 1st cache line + movq 0(%rax,$idx,8),%xmm4 # set of masks denoting which + movq 8(%rax,$idx,8),%xmm5 # cache line contains element + movq 16(%rax,$idx,8),%xmm6 # denoted by 7th argument + movq 24(%rax,$idx,8),%xmm7 + jmp .Lgather +.align 16 +.Lgather: + movq `0*$STRIDE/4-96`($tbl),%xmm0 + movq `1*$STRIDE/4-96`($tbl),%xmm1 + pand %xmm4,%xmm0 + movq `2*$STRIDE/4-96`($tbl),%xmm2 + pand %xmm5,%xmm1 + movq `3*$STRIDE/4-96`($tbl),%xmm3 + pand %xmm6,%xmm2 + por %xmm1,%xmm0 + pand %xmm7,%xmm3 + por %xmm2,%xmm0 + lea $STRIDE($tbl),$tbl + por %xmm3,%xmm0 + + movq %xmm0,($out) # m0=bp[0] + lea 8($out),$out + sub \$1,$num + jnz .Lgather +___ +$code.=<<___ if ($win64); + movaps %xmm6,(%rsp) + movaps %xmm7,0x10(%rsp) + lea 0x28(%rsp),%rsp +___ +$code.=<<___; + ret +.LSEH_end_bn_gather5: +.size bn_gather5,.-bn_gather5 ___ } $code.=<<___; @@ -980,6 +1039,10 @@ mul_handler: .rva .LSEH_end_bn_mul4x_mont_gather5 .rva .LSEH_info_bn_mul4x_mont_gather5 + .rva .LSEH_begin_bn_gather5 + .rva .LSEH_end_bn_gather5 + .rva .LSEH_info_bn_gather5 + .section .xdata .align 8 .LSEH_info_bn_mul_mont_gather5: @@ -992,6 +1055,12 @@ mul_handler: .rva mul_handler .rva .Lmul4x_alloca,.Lmul4x_body,.Lmul4x_epilogue # HandlerData[] .align 8 +.LSEH_info_bn_gather5: + .byte 0x01,0x0d,0x05,0x00 + .byte 0x0d,0x78,0x01,0x00 #movaps 0x10(rsp),xmm7 + .byte 0x08,0x68,0x00,0x00 #movaps (rsp),xmm6 + .byte 0x04,0x42,0x00,0x00 #sub rsp,0x28 +.align 8 ___ } diff --git a/crypto/bn/bn_exp.c b/crypto/bn/bn_exp.c index c69cd2c2f0..5c492365f3 100644 --- a/crypto/bn/bn_exp.c +++ b/crypto/bn/bn_exp.c @@ -535,23 +535,17 @@ err: * 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; } @@ -587,14 +581,13 @@ int BN_mod_exp_mont_consttime(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p, { int i,bits,ret=0,window,wvalue; int top; - BIGNUM *r; BN_MONT_CTX *mont=NULL; int numPowers; unsigned char *powerbufFree=NULL; int powerbufLen = 0; unsigned char *powerbuf=NULL; - BIGNUM computeTemp, *am=NULL; + BIGNUM tmp, am; bn_check_top(a); bn_check_top(p); @@ -614,10 +607,7 @@ 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. @@ -635,25 +625,13 @@ int BN_mod_exp_mont_consttime(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p, #if defined(OPENSSL_BN_ASM_MONT5) if (window==6 && bits<=1024) window=5; /* ~5% improvement of 2048-bit RSA sign */ #endif - /* 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; /* Allocate a buffer large enough to hold all of the pre-computed - * powers of a, plus computeTemp. + * powers of am, am itself and tmp. */ numPowers = 1 << window; powerbufLen = sizeof(m->d[0])*(top*numPowers + - (top>numPowers?top:numPowers)); + ((2*top)>numPowers?(2*top):numPowers)); #ifdef alloca if (powerbufLen < 3072) powerbufFree = alloca(powerbufLen+MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH); @@ -670,28 +648,31 @@ int BN_mod_exp_mont_consttime(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p, powerbufFree = NULL; #endif - computeTemp.d = (BN_ULONG *)(powerbuf + sizeof(m->d[0])*top*numPowers); - computeTemp.top = computeTemp.dmax = top; - computeTemp.neg = 0; - computeTemp.flags = BN_FLG_STATIC_DATA; + /* 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; - /* 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; - - /* Initialize computeTemp as a^1 with montgomery precalcs */ - am = BN_CTX_get(ctx); - if (am==NULL) goto err; + /* prepare a^0 in Montgomery domain */ +#if 1 + if (!BN_to_montgomery(&tmp,BN_value_one(),mont,ctx)) goto err; +#else + tmp.d[0] = (0-m->d[0])&BN_MASK2; /* 2^(top*BN_BITS2) - m */ + for (i=1;id[i])&BN_MASK2; + tmp.top = top; +#endif + /* prepare a^1 in Montgomery domain */ if (a->neg || BN_ucmp(a,m) >= 0) { - if (!BN_mod(am,a,m,ctx)) goto err; - if (!BN_to_montgomery(am,am,mont,ctx)) goto err; + 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 (!BN_copy(&computeTemp, am)) goto err; + else if (!BN_to_montgomery(&am,a,mont,ctx)) goto err; #if defined(OPENSSL_BN_ASM_MONT5) /* This optimization uses ideas from http://eprint.iacr.org/2011/239, @@ -707,95 +688,83 @@ int BN_mod_exp_mont_consttime(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p, 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); - BN_ULONG *acc, *np=mont->N.d, *n0=mont->n0; + BN_ULONG *np=mont->N.d, *n0=mont->n0; - bn_scatter5(r->d,r->top,powerbuf,0); - bn_scatter5(am->d,am->top,powerbuf,1); + 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); - acc = computeTemp.d; - /* bn_mul_mont() and bn_mul_mont_gather5() assume fixed length inputs. - * Pad the inputs with zeroes. - */ - if (bn_wexpand(am,top)==NULL || bn_wexpand(r,top)==NULL || - bn_wexpand(&computeTemp,top)==NULL) - goto err; - for (i = am->top; i < top; ++i) - { - am->d[i] = 0; - } - for (i = computeTemp.top; i < top; ++i) - { - computeTemp.d[i] = 0; - } - for (i = r->top; i < top; ++i) - { - r->d[i] = 0; - } #if 0 - for (i=2; i<32; i++) + for (i=3; i<32; i++) { - bn_mul_mont_gather5(acc,am->d,powerbuf,np,n0,top,i-1); - bn_scatter5(acc,top,powerbuf,i); + /* Calculate a^i = a^(i-1) * a */ + bn_mul_mont_gather5(tmp.d,am.d,powerbuf,np,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=2; i<32; i*=2) + for (i=4; i<32; i*=2) { - bn_mul_mont(acc,acc,acc,np,n0,top); - bn_scatter5(acc,top,powerbuf,i); + 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(acc,am->d,powerbuf,np,n0,top,i-1); - bn_scatter5(acc,top,powerbuf,i); + bn_mul_mont_gather5(tmp.d,am.d,powerbuf,np,n0,top,i-1); + bn_scatter5(tmp.d,top,powerbuf,i); for (j=2*i; j<32; j*=2) { - bn_mul_mont(acc,acc,acc,np,n0,top); - bn_scatter5(acc,top,powerbuf,j); + 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(acc,am->d,powerbuf,np,n0,top,i-1); - bn_scatter5(acc,top,powerbuf,i); - bn_mul_mont(acc,acc,acc,np,n0,top); - bn_scatter5(acc,top,powerbuf,2*i); + bn_mul_mont_gather5(tmp.d,am.d,powerbuf,np,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(acc,am->d,powerbuf,np,n0,top,i-1); - bn_scatter5(acc,top,powerbuf,i); + bn_mul_mont_gather5(tmp.d,am.d,powerbuf,np,n0,top,i-1); + bn_scatter5(tmp.d,top,powerbuf,i); } #endif - acc = r->d; + 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. */ - bits--; while (bits >= 0) { for (wvalue=0, i=0; i<5; i++,bits--) wvalue = (wvalue<<1)+BN_is_bit_set(p,bits); - bn_mul_mont(acc,acc,acc,np,n0,top); - bn_mul_mont(acc,acc,acc,np,n0,top); - bn_mul_mont(acc,acc,acc,np,n0,top); - bn_mul_mont(acc,acc,acc,np,n0,top); - bn_mul_mont(acc,acc,acc,np,n0,top); - bn_mul_mont_gather5(acc,acc,powerbuf,np,n0,top,wvalue); + 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); } - r->top=top; - bn_correct_top(r); + tmp.top=top; + bn_correct_top(&tmp); } else #endif { - if (!MOD_EXP_CTIME_COPY_TO_PREBUF(r, top, powerbuf, 0, numPowers)) goto err; - if (!MOD_EXP_CTIME_COPY_TO_PREBUF(am, top, powerbuf, 1, numPowers)) goto err; + 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) @@ -804,19 +773,25 @@ int BN_mod_exp_mont_consttime(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p, */ if (window > 1) { - for (i=2; 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 */ @@ -824,20 +799,20 @@ int BN_mod_exp_mont_consttime(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p, /* Scan the window, squaring the result as we go */ for (i=0; i