8  * Optimization for constant divisors on 32-bit machines:
53 #elif BITS_PER_LONG == 32
83 	/* compute m = ((p << 64) + b - 1) / b */			\
90 	/* test our ___m with res = m * x / (p << 64) */		\
92 	___t = (___m & 0xffffffff) * (___x >> 32) + (___res >> 32);	\
93 	___res = (___m >> 32) * (___x >> 32) + (___t >> 32);		\
94 	___t = (___m >> 32) * (___x & 0xffffffff) + (___t & 0xffffffff);\
95 	___res = (___res + (___t >> 32)) / ___p;			\
102 		 * additional bit to represent m which would overflow	\
105 		 * Instead we do m = p / b and n / b = (n * m + m) / p.	\
108 		/* Compute m = (p << 64) / b */				\
113 	/* Reduce m / p to help avoid overflow handling later. */	\
118 	 * Perform (m_bias + m * n) / (1 << 64).			\
130  * Prototype: uint64_t __arch_xprod_64(const uint64_t m, uint64_t n, bool bias)
131  * Semantic:  retval = ((bias ? m : 0) + m * n) >> 64
142 uint64_t __arch_xprod_64(const uint64_t m, uint64_t n, bool bias)  in __arch_xprod_64()  argument
144 	uint32_t m_lo = m;  in __arch_xprod_64()
145 	uint32_t m_hi = m >> 32;  in __arch_xprod_64()
147 	uint32_t n_hi = n >> 32;  in __arch_xprod_64()
151 	bool no_ovf = __builtin_constant_p(m) &&  in __arch_xprod_64()
152 		      ((m >> 32) + (m & 0xffffffff) < 0x100000000);  in __arch_xprod_64()
155 		x = (uint64_t)m_lo * n_lo + (bias ? m : 0);  in __arch_xprod_64()
156 		x >>= 32;  in __arch_xprod_64()
159 		x >>= 32;  in __arch_xprod_64()
163 		y = (uint64_t)m_lo * n_hi + (uint32_t)(x >> 32) + (bias ? m_hi : 0);  in __arch_xprod_64()
164 		x = (uint64_t)m_hi * n_hi + (uint32_t)(y >> 32);  in __arch_xprod_64()
166 		x += (uint32_t)(y >> 32);  in __arch_xprod_64()
192 		/* the remainder can be computed with 32-bit regs */ \
195 	} else if (likely(((n) >> 32) == 0)) {		\