From mboxrd@z Thu Jan 1 00:00:00 1970 Return-path: Received: from metis.ext.pengutronix.de ([2001:67c:670:201:290:27ff:fe1d:cc33]) by merlin.infradead.org with esmtps (Exim 4.92.3 #3 (Red Hat Linux)) id 1lCfm6-00063n-OJ for barebox@lists.infradead.org; Thu, 18 Feb 2021 09:39:08 +0000 From: Ahmad Fatoum Date: Thu, 18 Feb 2021 10:39:00 +0100 Message-Id: <20210218093902.15354-3-a.fatoum@pengutronix.de> In-Reply-To: <20210218093902.15354-1-a.fatoum@pengutronix.de> References: <20210218093902.15354-1-a.fatoum@pengutronix.de> MIME-Version: 1.0 List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Sender: "barebox" Errors-To: barebox-bounces+u.kleine-koenig=pengutronix.de@lists.infradead.org Subject: [PATCH v3 3/5] include: asm-generic: don't do 64-bit soft division on 64-bit platforms To: barebox@lists.infradead.org Cc: Ahmad Fatoum barebox implements do_div() as soft division unconditionally, even on 64-bit platforms that could use hardware 64-bit division directly. Import the whole Linux asm-generic/div64.h header to avoid this. This also has potential positive effect on 32-bit platforms: 64-bit division with constant divisors can now be optimized into multiplications and shifts at compile time. Signed-off-by: Ahmad Fatoum --- v2 -> v3: - no change v1 -> v2: - no change --- include/asm-generic/div64.h | 216 +++++++++++++++++++++++++++++++++++- 1 file changed, 215 insertions(+), 1 deletion(-) diff --git a/include/asm-generic/div64.h b/include/asm-generic/div64.h index 2e0ba8389e11..a3b98c86f077 100644 --- a/include/asm-generic/div64.h +++ b/include/asm-generic/div64.h @@ -1,9 +1,13 @@ +/* SPDX-License-Identifier: GPL-2.0 */ #ifndef _ASM_GENERIC_DIV64_H #define _ASM_GENERIC_DIV64_H /* * Copyright (C) 2003 Bernardo Innocenti * Based on former asm-ppc/div64.h and asm-m68knommu/div64.h * + * Optimization for constant divisors on 32-bit machines: + * Copyright (C) 2006-2015 Nicolas Pitre + * * The semantics of do_div() are: * * uint32_t do_div(uint64_t *n, uint32_t base) @@ -18,8 +22,200 @@ */ #include +#include + +#if BITS_PER_LONG == 64 + +/** + * do_div - returns 2 values: calculate remainder and update new dividend + * @n: uint64_t dividend (will be updated) + * @base: uint32_t divisor + * + * Summary: + * ``uint32_t remainder = n % base;`` + * ``n = n / base;`` + * + * Return: (uint32_t)remainder + * + * NOTE: macro parameter @n is evaluated multiple times, + * beware of side effects! + */ +# define do_div(n,base) ({ \ + uint32_t __base = (base); \ + uint32_t __rem; \ + __rem = ((uint64_t)(n)) % __base; \ + (n) = ((uint64_t)(n)) / __base; \ + __rem; \ + }) + +#elif BITS_PER_LONG == 32 + +#include + +/* + * If the divisor happens to be constant, we determine the appropriate + * inverse at compile time to turn the division into a few inline + * multiplications which ought to be much faster. And yet only if compiling + * with a sufficiently recent gcc version to perform proper 64-bit constant + * propagation. + * + * (It is unfortunate that gcc doesn't perform all this internally.) + */ +#ifndef __div64_const32_is_OK +#define __div64_const32_is_OK (__GNUC__ >= 4) +#endif + +#define __div64_const32(n, ___b) \ +({ \ + /* \ + * Multiplication by reciprocal of b: n / b = n * (p / b) / p \ + * \ + * We rely on the fact that most of this code gets optimized \ + * away at compile time due to constant propagation and only \ + * a few multiplication instructions should remain. \ + * Hence this monstrous macro (static inline doesn't always \ + * do the trick here). \ + */ \ + uint64_t ___res, ___x, ___t, ___m, ___n = (n); \ + uint32_t ___p, ___bias; \ + \ + /* determine MSB of b */ \ + ___p = 1 << ilog2(___b); \ + \ + /* compute m = ((p << 64) + b - 1) / b */ \ + ___m = (~0ULL / ___b) * ___p; \ + ___m += (((~0ULL % ___b + 1) * ___p) + ___b - 1) / ___b; \ + \ + /* one less than the dividend with highest result */ \ + ___x = ~0ULL / ___b * ___b - 1; \ + \ + /* test our ___m with res = m * x / (p << 64) */ \ + ___res = ((___m & 0xffffffff) * (___x & 0xffffffff)) >> 32; \ + ___t = ___res += (___m & 0xffffffff) * (___x >> 32); \ + ___res += (___x & 0xffffffff) * (___m >> 32); \ + ___t = (___res < ___t) ? (1ULL << 32) : 0; \ + ___res = (___res >> 32) + ___t; \ + ___res += (___m >> 32) * (___x >> 32); \ + ___res /= ___p; \ + \ + /* Now sanitize and optimize what we've got. */ \ + if (~0ULL % (___b / (___b & -___b)) == 0) { \ + /* special case, can be simplified to ... */ \ + ___n /= (___b & -___b); \ + ___m = ~0ULL / (___b / (___b & -___b)); \ + ___p = 1; \ + ___bias = 1; \ + } else if (___res != ___x / ___b) { \ + /* \ + * We can't get away without a bias to compensate \ + * for bit truncation errors. To avoid it we'd need an \ + * additional bit to represent m which would overflow \ + * a 64-bit variable. \ + * \ + * Instead we do m = p / b and n / b = (n * m + m) / p. \ + */ \ + ___bias = 1; \ + /* Compute m = (p << 64) / b */ \ + ___m = (~0ULL / ___b) * ___p; \ + ___m += ((~0ULL % ___b + 1) * ___p) / ___b; \ + } else { \ + /* \ + * Reduce m / p, and try to clear bit 31 of m when \ + * possible, otherwise that'll need extra overflow \ + * handling later. \ + */ \ + uint32_t ___bits = -(___m & -___m); \ + ___bits |= ___m >> 32; \ + ___bits = (~___bits) << 1; \ + /* \ + * If ___bits == 0 then setting bit 31 is unavoidable. \ + * Simply apply the maximum possible reduction in that \ + * case. Otherwise the MSB of ___bits indicates the \ + * best reduction we should apply. \ + */ \ + if (!___bits) { \ + ___p /= (___m & -___m); \ + ___m /= (___m & -___m); \ + } else { \ + ___p >>= ilog2(___bits); \ + ___m >>= ilog2(___bits); \ + } \ + /* No bias needed. */ \ + ___bias = 0; \ + } \ + \ + /* \ + * Now we have a combination of 2 conditions: \ + * \ + * 1) whether or not we need to apply a bias, and \ + * \ + * 2) whether or not there might be an overflow in the cross \ + * product determined by (___m & ((1 << 63) | (1 << 31))). \ + * \ + * Select the best way to do (m_bias + m * n) / (1 << 64). \ + * From now on there will be actual runtime code generated. \ + */ \ + ___res = __arch_xprod_64(___m, ___n, ___bias); \ + \ + ___res /= ___p; \ +}) + +#ifndef __arch_xprod_64 +/* + * Default C implementation for __arch_xprod_64() + * + * Prototype: uint64_t __arch_xprod_64(const uint64_t m, uint64_t n, bool bias) + * Semantic: retval = ((bias ? m : 0) + m * n) >> 64 + * + * The product is a 128-bit value, scaled down to 64 bits. + * Assuming constant propagation to optimize away unused conditional code. + * Architectures may provide their own optimized assembly implementation. + */ +static inline uint64_t __arch_xprod_64(const uint64_t m, uint64_t n, bool bias) +{ + uint32_t m_lo = m; + uint32_t m_hi = m >> 32; + uint32_t n_lo = n; + uint32_t n_hi = n >> 32; + uint64_t res; + uint32_t res_lo, res_hi, tmp; + + if (!bias) { + res = ((uint64_t)m_lo * n_lo) >> 32; + } else if (!(m & ((1ULL << 63) | (1ULL << 31)))) { + /* there can't be any overflow here */ + res = (m + (uint64_t)m_lo * n_lo) >> 32; + } else { + res = m + (uint64_t)m_lo * n_lo; + res_lo = res >> 32; + res_hi = (res_lo < m_hi); + res = res_lo | ((uint64_t)res_hi << 32); + } + + if (!(m & ((1ULL << 63) | (1ULL << 31)))) { + /* there can't be any overflow here */ + res += (uint64_t)m_lo * n_hi; + res += (uint64_t)m_hi * n_lo; + res >>= 32; + } else { + res += (uint64_t)m_lo * n_hi; + tmp = res >> 32; + res += (uint64_t)m_hi * n_lo; + res_lo = res >> 32; + res_hi = (res_lo < tmp); + res = res_lo | ((uint64_t)res_hi << 32); + } + + res += (uint64_t)m_hi * n_hi; + + return res; +} +#endif + +#ifndef __div64_32 extern uint32_t __div64_32(uint64_t *dividend, uint32_t divisor); +#endif /* The unnecessary pointer compare is there * to check for type safety (n must be 64bit) @@ -28,7 +224,19 @@ extern uint32_t __div64_32(uint64_t *dividend, uint32_t divisor); uint32_t __base = (base); \ uint32_t __rem; \ (void)(((typeof((n)) *)0) == ((uint64_t *)0)); \ - if (((n) >> 32) == 0) { \ + if (__builtin_constant_p(__base) && \ + is_power_of_2(__base)) { \ + __rem = (n) & (__base - 1); \ + (n) >>= ilog2(__base); \ + } else if (__div64_const32_is_OK && \ + __builtin_constant_p(__base) && \ + __base != 0) { \ + uint32_t __res_lo, __n_lo = (n); \ + (n) = __div64_const32(n, __base); \ + /* the remainder can be computed with 32-bit regs */ \ + __res_lo = (n); \ + __rem = __n_lo - __res_lo * __base; \ + } else if (likely(((n) >> 32) == 0)) { \ __rem = (uint32_t)(n) % __base; \ (n) = (uint32_t)(n) / __base; \ } else \ @@ -36,4 +244,10 @@ extern uint32_t __div64_32(uint64_t *dividend, uint32_t divisor); __rem; \ }) +#else /* BITS_PER_LONG == ?? */ + +# error do_div() does not yet support the C64 + +#endif /* BITS_PER_LONG */ + #endif /* _ASM_GENERIC_DIV64_H */ -- 2.29.2 _______________________________________________ barebox mailing list barebox@lists.infradead.org http://lists.infradead.org/mailman/listinfo/barebox