/** @file ***************************************************************************** Assembly code snippets for F[p] finite field arithmetic, used by fp.tcc . Specific to x86-64, and used only if USE_ASM is defined. On other architectures or without USE_ASM, fp.tcc uses a portable C++ implementation instead. ***************************************************************************** * @author This file is part of libsnark, developed by SCIPR Lab * and contributors (see AUTHORS). * @copyright MIT license (see LICENSE file) *****************************************************************************/ #ifndef FP_AUX_TCC_ #define FP_AUX_TCC_ namespace libsnark { #define STR_HELPER(x) #x #define STR(x) STR_HELPER(x) /* addq is faster than adcq, even if preceded by clc */ #define ADD_FIRSTADD \ "movq (%[B]), %%rax \n\t" \ "addq %%rax, (%[A]) \n\t" #define ADD_NEXTADD(ofs) \ "movq " STR(ofs) "(%[B]), %%rax \n\t" \ "adcq %%rax, " STR(ofs) "(%[A]) \n\t" #define ADD_CMP(ofs) \ "movq " STR(ofs) "(%[mod]), %%rax \n\t" \ "cmpq %%rax, " STR(ofs) "(%[A]) \n\t" \ "jb done%= \n\t" \ "ja subtract%= \n\t" #define ADD_FIRSTSUB \ "movq (%[mod]), %%rax \n\t" \ "subq %%rax, (%[A]) \n\t" #define ADD_FIRSTSUB \ "movq (%[mod]), %%rax \n\t" \ "subq %%rax, (%[A]) \n\t" #define ADD_NEXTSUB(ofs) \ "movq " STR(ofs) "(%[mod]), %%rax \n\t" \ "sbbq %%rax, " STR(ofs) "(%[A]) \n\t" #define SUB_FIRSTSUB \ "movq (%[B]), %%rax\n\t" \ "subq %%rax, (%[A])\n\t" #define SUB_NEXTSUB(ofs) \ "movq " STR(ofs) "(%[B]), %%rax\n\t" \ "sbbq %%rax, " STR(ofs) "(%[A])\n\t" #define SUB_FIRSTADD \ "movq (%[mod]), %%rax\n\t" \ "addq %%rax, (%[A])\n\t" #define SUB_NEXTADD(ofs) \ "movq " STR(ofs) "(%[mod]), %%rax\n\t" \ "adcq %%rax, " STR(ofs) "(%[A])\n\t" #define MONT_CMP(ofs) \ "movq " STR(ofs) "(%[M]), %%rax \n\t" \ "cmpq %%rax, " STR(ofs) "(%[tmp]) \n\t" \ "jb done%= \n\t" \ "ja subtract%= \n\t" #define MONT_FIRSTSUB \ "movq (%[M]), %%rax \n\t" \ "subq %%rax, (%[tmp]) \n\t" #define MONT_NEXTSUB(ofs) \ "movq " STR(ofs) "(%[M]), %%rax \n\t" \ "sbbq %%rax, " STR(ofs) "(%[tmp]) \n\t" /* The x86-64 Montgomery multiplication here is similar to Algorithm 2 (CIOS method) in http://eprint.iacr.org/2012/140.pdf and the PowerPC pseudocode of gmp-ecm library (c) Paul Zimmermann and Alexander Kruppa (see comments on top of powerpc64/mulredc.m4). */ #define MONT_PRECOMPUTE \ "xorq %[cy], %[cy] \n\t" \ "movq 0(%[A]), %%rax \n\t" \ "mulq 0(%[B]) \n\t" \ "movq %%rax, %[T0] \n\t" \ "movq %%rdx, %[T1] # T1:T0 <- A[0] * B[0] \n\t" \ "mulq %[inv] \n\t" \ "movq %%rax, %[u] # u <- T0 * inv \n\t" \ "mulq 0(%[M]) \n\t" \ "addq %[T0], %%rax \n\t" \ "adcq %%rdx, %[T1] \n\t" \ "adcq $0, %[cy] # cy:T1 <- (M[0]*u + T1 * b + T0) / b\n\t" #define MONT_FIRSTITER(j) \ "xorq %[T0], %[T0] \n\t" \ "movq 0(%[A]), %%rax \n\t" \ "mulq " STR((j*8)) "(%[B]) \n\t" \ "addq %[T1], %%rax \n\t" \ "movq %%rax, " STR(((j-1)*8)) "(%[tmp]) \n\t" \ "adcq $0, %%rdx \n\t" \ "movq %%rdx, %[T1] # now T1:tmp[j-1] <-- X[0] * Y[j] + T1\n\t" \ "movq " STR((j*8)) "(%[M]), %%rax \n\t" \ "mulq %[u] \n\t" \ "addq %%rax, " STR(((j-1)*8)) "(%[tmp]) \n\t" \ "adcq %[cy], %%rdx \n\t" \ "adcq $0, %[T0] \n\t" \ "xorq %[cy], %[cy] \n\t" \ "addq %%rdx, %[T1] \n\t" \ "adcq %[T0], %[cy] # cy:T1:tmp[j-1] <---- (X[0] * Y[j] + T1) + (M[j] * u + cy * b) \n\t" #define MONT_ITERFIRST(i) \ "xorq %[cy], %[cy] \n\t" \ "movq " STR((i*8)) "(%[A]), %%rax \n\t" \ "mulq 0(%[B]) \n\t" \ "addq 0(%[tmp]), %%rax \n\t" \ "adcq 8(%[tmp]), %%rdx \n\t" \ "adcq $0, %[cy] \n\t" \ "movq %%rax, %[T0] \n\t" \ "movq %%rdx, %[T1] # cy:T1:T0 <- A[i] * B[0] + tmp[1] * b + tmp[0]\n\t" \ "mulq %[inv] \n\t" \ "movq %%rax, %[u] # u <- T0 * inv\n\t" \ "mulq 0(%[M]) \n\t" \ "addq %[T0], %%rax \n\t" \ "adcq %%rdx, %[T1] \n\t" \ "adcq $0, %[cy] # cy:T1 <- (M[0]*u + cy * b * b + T1 * b + T0) / b\n\t" #define MONT_ITERITER(i, j) \ "xorq %[T0], %[T0] \n\t" \ "movq " STR((i*8)) "(%[A]), %%rax \n\t" \ "mulq " STR((j*8)) "(%[B]) \n\t" \ "addq %[T1], %%rax \n\t" \ "movq %%rax, " STR(((j-1)*8)) "(%[tmp]) \n\t" \ "adcq $0, %%rdx \n\t" \ "movq %%rdx, %[T1] # now T1:tmp[j-1] <-- X[i] * Y[j] + T1 \n\t" \ "movq " STR((j*8)) "(%[M]), %%rax \n\t" \ "mulq %[u] \n\t" \ "addq %%rax, " STR(((j-1)*8)) "(%[tmp]) \n\t" \ "adcq %[cy], %%rdx \n\t" \ "adcq $0, %[T0] \n\t" \ "xorq %[cy], %[cy] \n\t" \ "addq %%rdx, %[T1] \n\t" \ "adcq %[T0], %[cy] # cy:T1:tmp[j-1] <-- (X[i] * Y[j] + T1) + M[j] * u + cy * b \n\t" \ "addq " STR(((j+1)*8)) "(%[tmp]), %[T1] \n\t" \ "adcq $0, %[cy] # cy:T1:tmp[j-1] <-- (X[i] * Y[j] + T1) + M[j] * u + (tmp[j+1] + cy) * b \n\t" #define MONT_FINALIZE(j) \ "movq %[T1], " STR((j*8)) "(%[tmp]) \n\t" \ "movq %[cy], " STR(((j+1)*8)) "(%[tmp]) \n\t" /* Comba multiplication and squaring routines are based on the public-domain tomsfastmath library by Tom St Denis Compared to the above, we save 5-20% of cycles by using careful register renaming to implement Comba forward operation. */ #define COMBA_3_BY_3_MUL(c0_, c1_, c2_, res_, A_, B_) \ asm volatile ( \ "movq 0(%[A]), %%rax \n\t" \ "mulq 0(%[B]) \n\t" \ "movq %%rax, 0(%[res]) \n\t" \ "movq %%rdx, %[c0] \n\t" \ \ "xorq %[c1], %[c1] \n\t" \ "movq 0(%[A]), %%rax \n\t" \ "mulq 8(%[B]) \n\t" \ "addq %%rax, %[c0] \n\t" \ "adcq %%rdx, %[c1] \n\t" \ \ "xorq %[c2], %[c2] \n\t" \ "movq 8(%[A]), %%rax \n\t" \ "mulq 0(%[B]) \n\t" \ "addq %%rax, %[c0] \n\t" \ "movq %[c0], 8(%[res]) \n\t" \ "adcq %%rdx, %[c1] \n\t" \ "adcq $0, %[c2] \n\t" \ \ "// register renaming (c1, c2, c0)\n\t" \ "xorq %[c0], %[c0] \n\t" \ "movq 0(%[A]), %%rax \n\t" \ "mulq 16(%[B]) \n\t" \ "addq %%rax, %[c1] \n\t" \ "adcq %%rdx, %[c2] \n\t" \ "adcq $0, %[c0] \n\t" \ \ "movq 8(%[A]), %%rax \n\t" \ "mulq 8(%[B]) \n\t" \ "addq %%rax, %[c1] \n\t" \ "adcq %%rdx, %[c2] \n\t" \ "adcq $0, %[c0] \n\t" \ \ "movq 16(%[A]), %%rax \n\t" \ "mulq 0(%[B]) \n\t" \ "addq %%rax, %[c1] \n\t" \ "movq %[c1], 16(%[res]) \n\t" \ "adcq %%rdx, %[c2] \n\t" \ "adcq $0, %[c0] \n\t" \ \ "// register renaming (c2, c0, c1)\n\t" \ "xorq %[c1], %[c1] \n\t" \ "movq 8(%[A]), %%rax \n\t" \ "mulq 16(%[B]) \n\t" \ "addq %%rax, %[c2] \n\t" \ "adcq %%rdx, %[c0] \n\t" \ "adcq $0, %[c1] \n\t" \ \ "movq 16(%[A]), %%rax \n\t" \ "mulq 8(%[B]) \n\t" \ "addq %%rax, %[c2] \n\t" \ "movq %[c2], 24(%[res]) \n\t" \ "adcq %%rdx, %[c0] \n\t" \ "adcq $0, %[c1] \n\t" \ \ "// register renaming (c0, c1, c2)\n\t" \ "xorq %[c2], %[c2] \n\t" \ "movq 16(%[A]), %%rax \n\t" \ "mulq 16(%[B]) \n\t" \ "addq %%rax, %[c0] \n\t" \ "movq %[c0], 32(%[res]) \n\t" \ "adcq %%rdx, %[c1] \n\t" \ "movq %[c1], 40(%[res]) \n\t" \ : [c0] "=&r" (c0_), [c1] "=&r" (c1_), [c2] "=&r" (c2_) \ : [res] "r" (res_), [A] "r" (A_), [B] "r" (B_) \ : "%rax", "%rdx", "cc", "memory") #define COMBA_3_BY_3_SQR(c0_, c1_, c2_, res_, A_) \ asm volatile ( \ "xorq %[c1], %[c1] \n\t" \ "xorq %[c2], %[c2] \n\t" \ "movq 0(%[A]), %%rax \n\t" \ "mulq %%rax \n\t" \ "movq %%rax, 0(%[res]) \n\t" \ "movq %%rdx, %[c0] \n\t" \ \ "movq 0(%[A]), %%rax \n\t" \ "mulq 8(%[A]) \n\t" \ "addq %%rax, %[c0] \n\t" \ "adcq %%rdx, %[c1] \n\t" \ "addq %%rax, %[c0] \n\t" \ "movq %[c0], 8(%[res]) \n\t" \ "adcq %%rdx, %[c1] \n\t" \ "adcq $0, %[c2] \n\t" \ \ "// register renaming (c1, c2, c0)\n\t" \ "movq 0(%[A]), %%rax \n\t" \ "xorq %[c0], %[c0] \n\t" \ "mulq 16(%[A]) \n\t" \ "addq %%rax, %[c1] \n\t" \ "adcq %%rdx, %[c2] \n\t" \ "adcq $0, %[c0] \n\t" \ "addq %%rax, %[c1] \n\t" \ "adcq %%rdx, %[c2] \n\t" \ "adcq $0, %[c0] \n\t" \ \ "movq 8(%[A]), %%rax \n\t" \ "mulq %%rax \n\t" \ "addq %%rax, %[c1] \n\t" \ "movq %[c1], 16(%[res]) \n\t" \ "adcq %%rdx, %[c2] \n\t" \ "adcq $0, %[c0] \n\t" \ \ "// register renaming (c2, c0, c1)\n\t" \ "movq 8(%[A]), %%rax \n\t" \ "xorq %[c1], %[c1] \n\t" \ "mulq 16(%[A]) \n\t" \ "addq %%rax, %[c2] \n\t" \ "adcq %%rdx, %[c0] \n\t" \ "adcq $0, %[c1] \n\t" \ "addq %%rax, %[c2] \n\t" \ "movq %[c2], 24(%[res]) \n\t" \ "adcq %%rdx, %[c0] \n\t" \ "adcq $0, %[c1] \n\t" \ \ "// register renaming (c0, c1, c2)\n\t" \ "movq 16(%[A]), %%rax \n\t" \ "mulq %%rax \n\t" \ "addq %%rax, %[c0] \n\t" \ "movq %[c0], 32(%[res]) \n\t" \ "adcq %%rdx, %[c1] \n\t" \ "movq %[c1], 40(%[res]) \n\t" \ \ : [c0] "=&r" (c0_), [c1] "=&r" (c1_), [c2] "=&r" (c2_) \ : [res] "r" (res_), [A] "r" (A_) \ : "%rax", "%rdx", "cc", "memory") /* The Montgomery reduction here is based on Algorithm 14.32 in Handbook of Applied Cryptography . */ #define REDUCE_6_LIMB_PRODUCT(k_, tmp1_, tmp2_, tmp3_, inv_, res_, mod_) \ __asm__ volatile \ ("///////////////////////////////////\n\t" \ "movq 0(%[res]), %%rax \n\t" \ "mulq %[modprime] \n\t" \ "movq %%rax, %[k] \n\t" \ \ "movq (%[mod]), %%rax \n\t" \ "mulq %[k] \n\t" \ "movq %%rax, %[tmp1] \n\t" \ "movq %%rdx, %[tmp2] \n\t" \ \ "xorq %[tmp3], %[tmp3] \n\t" \ "movq 8(%[mod]), %%rax \n\t" \ "mulq %[k] \n\t" \ "addq %[tmp1], 0(%[res]) \n\t" \ "adcq %%rax, %[tmp2] \n\t" \ "adcq %%rdx, %[tmp3] \n\t" \ \ "xorq %[tmp1], %[tmp1] \n\t" \ "movq 16(%[mod]), %%rax \n\t" \ "mulq %[k] \n\t" \ "addq %[tmp2], 8(%[res]) \n\t" \ "adcq %%rax, %[tmp3] \n\t" \ "adcq %%rdx, %[tmp1] \n\t" \ \ "addq %[tmp3], 16(%[res]) \n\t" \ "adcq %[tmp1], 24(%[res]) \n\t" \ "adcq $0, 32(%[res]) \n\t" \ "adcq $0, 40(%[res]) \n\t" \ \ "///////////////////////////////////\n\t" \ "movq 8(%[res]), %%rax \n\t" \ "mulq %[modprime] \n\t" \ "movq %%rax, %[k] \n\t" \ \ "movq (%[mod]), %%rax \n\t" \ "mulq %[k] \n\t" \ "movq %%rax, %[tmp1] \n\t" \ "movq %%rdx, %[tmp2] \n\t" \ \ "xorq %[tmp3], %[tmp3] \n\t" \ "movq 8(%[mod]), %%rax \n\t" \ "mulq %[k] \n\t" \ "addq %[tmp1], 8(%[res]) \n\t" \ "adcq %%rax, %[tmp2] \n\t" \ "adcq %%rdx, %[tmp3] \n\t" \ \ "xorq %[tmp1], %[tmp1] \n\t" \ "movq 16(%[mod]), %%rax \n\t" \ "mulq %[k] \n\t" \ "addq %[tmp2], 16(%[res]) \n\t" \ "adcq %%rax, %[tmp3] \n\t" \ "adcq %%rdx, %[tmp1] \n\t" \ \ "addq %[tmp3], 24(%[res]) \n\t" \ "adcq %[tmp1], 32(%[res]) \n\t" \ "adcq $0, 40(%[res]) \n\t" \ \ "///////////////////////////////////\n\t" \ "movq 16(%[res]), %%rax \n\t" \ "mulq %[modprime] \n\t" \ "movq %%rax, %[k] \n\t" \ \ "movq (%[mod]), %%rax \n\t" \ "mulq %[k] \n\t" \ "movq %%rax, %[tmp1] \n\t" \ "movq %%rdx, %[tmp2] \n\t" \ \ "xorq %[tmp3], %[tmp3] \n\t" \ "movq 8(%[mod]), %%rax \n\t" \ "mulq %[k] \n\t" \ "addq %[tmp1], 16(%[res]) \n\t" \ "adcq %%rax, %[tmp2] \n\t" \ "adcq %%rdx, %[tmp3] \n\t" \ \ "xorq %[tmp1], %[tmp1] \n\t" \ "movq 16(%[mod]), %%rax \n\t" \ "mulq %[k] \n\t" \ "addq %[tmp2], 24(%[res]) \n\t" \ "adcq %%rax, %[tmp3] \n\t" \ "adcq %%rdx, %[tmp1] \n\t" \ \ "addq %[tmp3], 32(%[res]) \n\t" \ "adcq %[tmp1], 40(%[res]) \n\t" \ : [k] "=&r" (k_), [tmp1] "=&r" (tmp1_), [tmp2] "=&r" (tmp2_), [tmp3] "=&r" (tmp3_) \ : [modprime] "r" (inv_), [res] "r" (res_), [mod] "r" (mod_) \ : "%rax", "%rdx", "cc", "memory") } // libsnark #endif // FP_AUX_TCC_