2
* 64x64->128 bit multiply operation
3
* (C) 2013,2015 Jack Lloyd
5
* Botan is released under the Simplified BSD License (see license.txt)
8
#ifndef BOTAN_UTIL_MUL128_H_
9
#define BOTAN_UTIL_MUL128_H_
11
#include <botan/types.h>
15
#if defined(__SIZEOF_INT128__) && defined(BOTAN_TARGET_CPU_HAS_NATIVE_64BIT) && !defined(__xlc__)
16
#define BOTAN_TARGET_HAS_NATIVE_UINT128
18
// Prefer TI mode over __int128 as GCC rejects the latter in pendantic mode
20
typedef unsigned int uint128_t __attribute__((mode(TI)));
22
typedef unsigned __int128 uint128_t;
28
#if defined(BOTAN_TARGET_HAS_NATIVE_UINT128)
30
#define BOTAN_FAST_64X64_MUL(a,b,lo,hi) \
32
const uint128_t r = static_cast<uint128_t>(a) * b; \
33
*hi = (r >> 64) & 0xFFFFFFFFFFFFFFFF; \
34
*lo = (r ) & 0xFFFFFFFFFFFFFFFF; \
37
#elif defined(BOTAN_BUILD_COMPILER_IS_MSVC) && defined(BOTAN_TARGET_CPU_HAS_NATIVE_64BIT)
40
#pragma intrinsic(_umul128)
42
#define BOTAN_FAST_64X64_MUL(a,b,lo,hi) \
43
do { *lo = _umul128(a, b, hi); } while(0)
45
#elif defined(BOTAN_USE_GCC_INLINE_ASM)
47
#if defined(BOTAN_TARGET_ARCH_IS_X86_64)
49
#define BOTAN_FAST_64X64_MUL(a,b,lo,hi) do { \
50
asm("mulq %3" : "=d" (*hi), "=a" (*lo) : "a" (a), "rm" (b) : "cc"); \
53
#elif defined(BOTAN_TARGET_ARCH_IS_ALPHA)
55
#define BOTAN_FAST_64X64_MUL(a,b,lo,hi) do { \
56
asm("umulh %1,%2,%0" : "=r" (*hi) : "r" (a), "r" (b)); \
60
#elif defined(BOTAN_TARGET_ARCH_IS_IA64)
62
#define BOTAN_FAST_64X64_MUL(a,b,lo,hi) do { \
63
asm("xmpy.hu %0=%1,%2" : "=f" (*hi) : "f" (a), "f" (b)); \
67
#elif defined(BOTAN_TARGET_ARCH_IS_PPC64)
69
#define BOTAN_FAST_64X64_MUL(a,b,lo,hi) do { \
70
asm("mulhdu %0,%1,%2" : "=r" (*hi) : "r" (a), "r" (b) : "cc"); \
81
* Perform a 64x64->128 bit multiplication
83
inline void mul64x64_128(uint64_t a, uint64_t b, uint64_t* lo, uint64_t* hi)
85
#if defined(BOTAN_FAST_64X64_MUL)
86
BOTAN_FAST_64X64_MUL(a, b, lo, hi);
90
* Do a 64x64->128 multiply using four 32x32->64 multiplies plus
91
* some adds and shifts. Last resort for CPUs like UltraSPARC (with
92
* 64-bit registers/ALU, but no 64x64->128 multiply) or 32-bit CPUs.
94
const size_t HWORD_BITS = 32;
95
const uint32_t HWORD_MASK = 0xFFFFFFFF;
97
const uint32_t a_hi = (a >> HWORD_BITS);
98
const uint32_t a_lo = (a & HWORD_MASK);
99
const uint32_t b_hi = (b >> HWORD_BITS);
100
const uint32_t b_lo = (b & HWORD_MASK);
102
uint64_t x0 = static_cast<uint64_t>(a_hi) * b_hi;
103
uint64_t x1 = static_cast<uint64_t>(a_lo) * b_hi;
104
uint64_t x2 = static_cast<uint64_t>(a_hi) * b_lo;
105
uint64_t x3 = static_cast<uint64_t>(a_lo) * b_lo;
107
// this cannot overflow as (2^32-1)^2 + 2^32-1 < 2^64-1
108
x2 += x3 >> HWORD_BITS;
110
// this one can overflow
113
// propagate the carry if any
114
x0 += static_cast<uint64_t>(static_cast<bool>(x2 < x1)) << HWORD_BITS;
116
*hi = x0 + (x2 >> HWORD_BITS);
117
*lo = ((x2 & HWORD_MASK) << HWORD_BITS) + (x3 & HWORD_MASK);