/* $OpenBSD: x86_64-gcc.c,v 1.7 2022/11/26 16:08:51 tb Exp $ */ #include "../bn_local.h" /* * x86_64 BIGNUM accelerator version 0.1, December 2002. * * Implemented by Andy Polyakov for the OpenSSL * project. * * Rights for redistribution and usage in source and binary forms are * granted according to the OpenSSL license. Warranty of any kind is * disclaimed. * * Q. Version 0.1? It doesn't sound like Andy, he used to assign real * versions, like 1.0... * A. Well, that's because this code is basically a quick-n-dirty * proof-of-concept hack. As you can see it's implemented with * inline assembler, which means that you're bound to GCC and that * there might be enough room for further improvement. * * Q. Why inline assembler? * A. x86_64 features own ABI which I'm not familiar with. This is * why I decided to let the compiler take care of subroutine * prologue/epilogue as well as register allocation. For reference. * Win64 implements different ABI for AMD64, different from Linux. * * Q. How much faster does it get? * A. 'apps/openssl speed rsa dsa' output with no-asm: * * sign verify sign/s verify/s * rsa 512 bits 0.0006s 0.0001s 1683.8 18456.2 * rsa 1024 bits 0.0028s 0.0002s 356.0 6407.0 * rsa 2048 bits 0.0172s 0.0005s 58.0 1957.8 * rsa 4096 bits 0.1155s 0.0018s 8.7 555.6 * sign verify sign/s verify/s * dsa 512 bits 0.0005s 0.0006s 2100.8 1768.3 * dsa 1024 bits 0.0014s 0.0018s 692.3 559.2 * dsa 2048 bits 0.0049s 0.0061s 204.7 165.0 * * 'apps/openssl speed rsa dsa' output with this module: * * sign verify sign/s verify/s * rsa 512 bits 0.0004s 0.0000s 2767.1 33297.9 * rsa 1024 bits 0.0012s 0.0001s 867.4 14674.7 * rsa 2048 bits 0.0061s 0.0002s 164.0 5270.0 * rsa 4096 bits 0.0384s 0.0006s 26.1 1650.8 * sign verify sign/s verify/s * dsa 512 bits 0.0002s 0.0003s 4442.2 3786.3 * dsa 1024 bits 0.0005s 0.0007s 1835.1 1497.4 * dsa 2048 bits 0.0016s 0.0020s 620.4 504.6 * * For the reference. IA-32 assembler implementation performs * very much like 64-bit code compiled with no-asm on the same * machine. */ #define BN_ULONG unsigned long #undef mul #undef mul_add #undef sqr /* * "m"(a), "+m"(r) is the way to favor DirectPath µ-code; * "g"(0) let the compiler to decide where does it * want to keep the value of zero; */ #define mul_add(r,a,word,carry) do { \ BN_ULONG high,low; \ asm ("mulq %3" \ : "=a"(low),"=d"(high) \ : "a"(word),"m"(a) \ : "cc"); \ asm ("addq %2,%0; adcq %3,%1" \ : "+r"(carry),"+d"(high)\ : "a"(low),"g"(0) \ : "cc"); \ asm ("addq %2,%0; adcq %3,%1" \ : "+m"(r),"+d"(high) \ : "r"(carry),"g"(0) \ : "cc"); \ carry=high; \ } while (0) #define mul(r,a,word,carry) do { \ BN_ULONG high,low; \ asm ("mulq %3" \ : "=a"(low),"=d"(high) \ : "a"(word),"g"(a) \ : "cc"); \ asm ("addq %2,%0; adcq %3,%1" \ : "+r"(carry),"+d"(high)\ : "a"(low),"g"(0) \ : "cc"); \ (r)=carry, carry=high; \ } while (0) #define sqr(r0,r1,a) \ asm ("mulq %2" \ : "=a"(r0),"=d"(r1) \ : "a"(a) \ : "cc"); BN_ULONG bn_mul_add_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w) { BN_ULONG c1=0; if (num <= 0) return(c1); while (num&~3) { mul_add(rp[0],ap[0],w,c1); mul_add(rp[1],ap[1],w,c1); mul_add(rp[2],ap[2],w,c1); mul_add(rp[3],ap[3],w,c1); ap+=4; rp+=4; num-=4; } if (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; } return(c1); } BN_ULONG bn_mul_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w) { BN_ULONG c1=0; if (num <= 0) return(c1); while (num&~3) { mul(rp[0],ap[0],w,c1); mul(rp[1],ap[1],w,c1); mul(rp[2],ap[2],w,c1); mul(rp[3],ap[3],w,c1); ap+=4; rp+=4; num-=4; } if (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); } return(c1); } void bn_sqr_words(BN_ULONG *r, const BN_ULONG *a, int n) { if (n <= 0) return; while (n&~3) { sqr(r[0],r[1],a[0]); sqr(r[2],r[3],a[1]); sqr(r[4],r[5],a[2]); sqr(r[6],r[7],a[3]); a+=4; r+=8; n-=4; } if (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]); } } BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d) { BN_ULONG ret,waste; asm ("divq %4" : "=a"(ret),"=d"(waste) : "a"(l),"d"(h),"g"(d) : "cc"); return ret; } BN_ULONG bn_add_words (BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,int n) { BN_ULONG ret=0,i=0; if (n <= 0) return 0; asm ( " subq %2,%2 \n" ".p2align 4 \n" "1: movq (%4,%2,8),%0 \n" " adcq (%5,%2,8),%0 \n" " movq %0,(%3,%2,8) \n" " leaq 1(%2),%2 \n" " loop 1b \n" " sbbq %0,%0 \n" : "=&a"(ret),"+c"(n),"=&r"(i) : "r"(rp),"r"(ap),"r"(bp) : "cc" ); return ret&1; } BN_ULONG bn_sub_words (BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,int n) { BN_ULONG ret=0,i=0; if (n <= 0) return 0; asm ( " subq %2,%2 \n" ".p2align 4 \n" "1: movq (%4,%2,8),%0 \n" " sbbq (%5,%2,8),%0 \n" " movq %0,(%3,%2,8) \n" " leaq 1(%2),%2 \n" " loop 1b \n" " sbbq %0,%0 \n" : "=&a"(ret),"+c"(n),"=&r"(i) : "r"(rp),"r"(ap),"r"(bp) : "cc" ); return ret&1; } /* mul_add_c(a,b,c0,c1,c2) -- c+=a*b for three word number c=(c2,c1,c0) */ /* mul_add_c2(a,b,c0,c1,c2) -- c+=2*a*b for three word number c=(c2,c1,c0) */ /* sqr_add_c(a,i,c0,c1,c2) -- c+=a[i]^2 for three word number c=(c2,c1,c0) */ /* sqr_add_c2(a,i,c0,c1,c2) -- c+=2*a[i]*a[j] for three word number c=(c2,c1,c0) */ /* * Keep in mind that carrying into high part of multiplication result * can not overflow, because it cannot be all-ones. */ #if 0 /* original macros are kept for reference purposes */ #define mul_add_c(a,b,c0,c1,c2) do { \ BN_ULONG ta = (a), tb = (b); \ BN_ULONG lo, hi; \ BN_UMULT_LOHI(lo,hi,ta,tb); \ c0 += lo; hi += (c0