]> git.sur5r.net Git - u-boot/commitdiff
lib: div64: sync with Linux
authorPeng Fan <peng.fan@nxp.com>
Mon, 10 Apr 2017 05:39:48 +0000 (13:39 +0800)
committerTom Rini <trini@konsulko.com>
Thu, 13 Apr 2017 13:41:10 +0000 (09:41 -0400)
Sync with Linux commit ad0376eb1483b ("Merge tag 'edac_for_4.11_2'").

Signed-off-by: Peng Fan <peng.fan@nxp.com>
Cc: Tom Rini <trini@konsulko.com>
include/div64.h
include/linux/math64.h
lib/div64.c

index d83314435d860d59ffb9d5a14284d70a3c3a4453..cecb2322bcbf0dbc8bc109782a9cc76745dbefa6 100644 (file)
@@ -4,13 +4,16 @@
  * Copyright (C) 2003 Bernardo Innocenti <bernie@develer.com>
  * 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)
  * {
- *     uint32_t remainder = *n % base;
- *     *n = *n / base;
- *     return remainder;
+ *     uint32_t remainder = *n % base;
+ *     *n = *n / base;
+ *     return remainder;
  * }
  *
  * NOTE: macro parameter n is evaluated multiple times,
  */
 
 #include <linux/types.h>
+#include <linux/compiler.h>
+
+#if BITS_PER_LONG == 64
+
+# 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 <linux/log2.h>
+
+/*
+ * 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, 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;
+               tmp = (res < m) ? (1ULL << 32) : 0;
+               res = (res >> 32) + tmp;
+       }
+
+       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 {
+               tmp = res += (uint64_t)m_lo * n_hi;
+               res += (uint64_t)m_hi * n_lo;
+               tmp = (res < tmp) ? (1ULL << 32) : 0;
+               res = (res >> 32) + tmp;
+       }
 
+       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)
@@ -28,14 +205,32 @@ 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                                          \
+       } else                                          \
                __rem = __div64_32(&(n), __base);       \
        __rem;                                          \
  })
 
+#else /* BITS_PER_LONG == ?? */
+
+# error do_div() does not yet support the C64
+
+#endif /* BITS_PER_LONG */
+
 /* Wrapper for do_div(). Doesn't modify dividend and returns
  * the result, not reminder.
  */
index 6d760d75c4f789e95776adc7733aba3fbe850829..08584c8f237f39640650c9b93ee794f6ba8f5b2d 100644 (file)
@@ -1,10 +1,15 @@
 #ifndef _LINUX_MATH64_H
 #define _LINUX_MATH64_H
 
+#include <div64.h>
+#include <linux/bitops.h>
 #include <linux/types.h>
 
 #if BITS_PER_LONG == 64
 
+#define div64_long(x, y) div64_s64((x), (y))
+#define div64_ul(x, y)   div64_u64((x), (y))
+
 /**
  * div_u64_rem - unsigned 64bit divide with 32bit divisor with remainder
  *
@@ -26,6 +31,15 @@ static inline s64 div_s64_rem(s64 dividend, s32 divisor, s32 *remainder)
        return dividend / divisor;
 }
 
+/**
+ * div64_u64_rem - unsigned 64bit divide with 64bit divisor and remainder
+ */
+static inline u64 div64_u64_rem(u64 dividend, u64 divisor, u64 *remainder)
+{
+       *remainder = dividend % divisor;
+       return dividend / divisor;
+}
+
 /**
  * div64_u64 - unsigned 64bit divide with 64bit divisor
  */
@@ -34,8 +48,19 @@ static inline u64 div64_u64(u64 dividend, u64 divisor)
        return dividend / divisor;
 }
 
+/**
+ * div64_s64 - signed 64bit divide with 64bit divisor
+ */
+static inline s64 div64_s64(s64 dividend, s64 divisor)
+{
+       return dividend / divisor;
+}
+
 #elif BITS_PER_LONG == 32
 
+#define div64_long(x, y) div_s64((x), (y))
+#define div64_ul(x, y)   div_u64((x), (y))
+
 #ifndef div_u64_rem
 static inline u64 div_u64_rem(u64 dividend, u32 divisor, u32 *remainder)
 {
@@ -48,10 +73,18 @@ static inline u64 div_u64_rem(u64 dividend, u32 divisor, u32 *remainder)
 extern s64 div_s64_rem(s64 dividend, s32 divisor, s32 *remainder);
 #endif
 
+#ifndef div64_u64_rem
+extern u64 div64_u64_rem(u64 dividend, u64 divisor, u64 *remainder);
+#endif
+
 #ifndef div64_u64
 extern u64 div64_u64(u64 dividend, u64 divisor);
 #endif
 
+#ifndef div64_s64
+extern s64 div64_s64(s64 dividend, s64 divisor);
+#endif
+
 #endif /* BITS_PER_LONG */
 
 /**
@@ -82,4 +115,143 @@ static inline s64 div_s64(s64 dividend, s32 divisor)
 
 u32 iter_div_u64_rem(u64 dividend, u32 divisor, u64 *remainder);
 
+static __always_inline u32
+__iter_div_u64_rem(u64 dividend, u32 divisor, u64 *remainder)
+{
+       u32 ret = 0;
+
+       while (dividend >= divisor) {
+               /* The following asm() prevents the compiler from
+                  optimising this loop into a modulo operation.  */
+               asm("" : "+rm"(dividend));
+
+               dividend -= divisor;
+               ret++;
+       }
+
+       *remainder = dividend;
+
+       return ret;
+}
+
+#ifndef mul_u32_u32
+/*
+ * Many a GCC version messes this up and generates a 64x64 mult :-(
+ */
+static inline u64 mul_u32_u32(u32 a, u32 b)
+{
+       return (u64)a * b;
+}
+#endif
+
+#if defined(CONFIG_ARCH_SUPPORTS_INT128) && defined(__SIZEOF_INT128__)
+
+#ifndef mul_u64_u32_shr
+static inline u64 mul_u64_u32_shr(u64 a, u32 mul, unsigned int shift)
+{
+       return (u64)(((unsigned __int128)a * mul) >> shift);
+}
+#endif /* mul_u64_u32_shr */
+
+#ifndef mul_u64_u64_shr
+static inline u64 mul_u64_u64_shr(u64 a, u64 mul, unsigned int shift)
+{
+       return (u64)(((unsigned __int128)a * mul) >> shift);
+}
+#endif /* mul_u64_u64_shr */
+
+#else
+
+#ifndef mul_u64_u32_shr
+static inline u64 mul_u64_u32_shr(u64 a, u32 mul, unsigned int shift)
+{
+       u32 ah, al;
+       u64 ret;
+
+       al = a;
+       ah = a >> 32;
+
+       ret = mul_u32_u32(al, mul) >> shift;
+       if (ah)
+               ret += mul_u32_u32(ah, mul) << (32 - shift);
+
+       return ret;
+}
+#endif /* mul_u64_u32_shr */
+
+#ifndef mul_u64_u64_shr
+static inline u64 mul_u64_u64_shr(u64 a, u64 b, unsigned int shift)
+{
+       union {
+               u64 ll;
+               struct {
+#ifdef __BIG_ENDIAN
+                       u32 high, low;
+#else
+                       u32 low, high;
+#endif
+               } l;
+       } rl, rm, rn, rh, a0, b0;
+       u64 c;
+
+       a0.ll = a;
+       b0.ll = b;
+
+       rl.ll = mul_u32_u32(a0.l.low, b0.l.low);
+       rm.ll = mul_u32_u32(a0.l.low, b0.l.high);
+       rn.ll = mul_u32_u32(a0.l.high, b0.l.low);
+       rh.ll = mul_u32_u32(a0.l.high, b0.l.high);
+
+       /*
+        * Each of these lines computes a 64-bit intermediate result into "c",
+        * starting at bits 32-95.  The low 32-bits go into the result of the
+        * multiplication, the high 32-bits are carried into the next step.
+        */
+       rl.l.high = c = (u64)rl.l.high + rm.l.low + rn.l.low;
+       rh.l.low = c = (c >> 32) + rm.l.high + rn.l.high + rh.l.low;
+       rh.l.high = (c >> 32) + rh.l.high;
+
+       /*
+        * The 128-bit result of the multiplication is in rl.ll and rh.ll,
+        * shift it right and throw away the high part of the result.
+        */
+       if (shift == 0)
+               return rl.ll;
+       if (shift < 64)
+               return (rl.ll >> shift) | (rh.ll << (64 - shift));
+       return rh.ll >> (shift & 63);
+}
+#endif /* mul_u64_u64_shr */
+
+#endif
+
+#ifndef mul_u64_u32_div
+static inline u64 mul_u64_u32_div(u64 a, u32 mul, u32 divisor)
+{
+       union {
+               u64 ll;
+               struct {
+#ifdef __BIG_ENDIAN
+                       u32 high, low;
+#else
+                       u32 low, high;
+#endif
+               } l;
+       } u, rl, rh;
+
+       u.ll = a;
+       rl.ll = mul_u32_u32(u.l.low, mul);
+       rh.ll = mul_u32_u32(u.l.high, mul) + rl.l.high;
+
+       /* Bits 32-63 of the result will be in rh.l.low. */
+       rl.l.high = do_div(rh.ll, divisor);
+
+       /* Bits 0-31 of the result will be in rl.l.low. */
+       do_div(rl.ll, divisor);
+
+       rl.l.high = rh.l.low;
+       return rl.ll;
+}
+#endif /* mul_u64_u32_div */
+
 #endif /* _LINUX_MATH64_H */
index 319fca50fa79e2a61b2bbb1b48f3dbecb339647e..206f582ca96c5a11df528cd7dfc7f16eb57145da 100644 (file)
  *
  * Code generated for this function might be very inefficient
  * for some CPUs. __div64_32() can be overridden by linking arch-specific
- * assembly versions such as arch/powerpc/lib/div64.S and arch/sh/lib/div64.S.
+ * assembly versions such as arch/ppc/lib/div64.S and arch/sh/lib/div64.S
+ * or by defining a preprocessor macro in arch/include/asm/div64.h.
  */
 
-#include <div64.h>
-#include <linux/types.h>
-#include <linux/compiler.h>
+#include <linux/compat.h>
+#include <linux/kernel.h>
+#include <linux/math64.h>
 
-uint32_t notrace __div64_32(uint64_t *n, uint32_t base)
+/* Not needed on 64bit architectures */
+#if BITS_PER_LONG == 32
+
+#ifndef __div64_32
+uint32_t __attribute__((weak)) __div64_32(uint64_t *n, uint32_t base)
 {
        uint64_t rem = *n;
        uint64_t b = base;
@@ -52,3 +57,129 @@ uint32_t notrace __div64_32(uint64_t *n, uint32_t base)
        *n = res;
        return rem;
 }
+EXPORT_SYMBOL(__div64_32);
+#endif
+
+#ifndef div_s64_rem
+s64 div_s64_rem(s64 dividend, s32 divisor, s32 *remainder)
+{
+       u64 quotient;
+
+       if (dividend < 0) {
+               quotient = div_u64_rem(-dividend, abs(divisor), (u32 *)remainder);
+               *remainder = -*remainder;
+               if (divisor > 0)
+                       quotient = -quotient;
+       } else {
+               quotient = div_u64_rem(dividend, abs(divisor), (u32 *)remainder);
+               if (divisor < 0)
+                       quotient = -quotient;
+       }
+       return quotient;
+}
+EXPORT_SYMBOL(div_s64_rem);
+#endif
+
+/**
+ * div64_u64_rem - unsigned 64bit divide with 64bit divisor and remainder
+ * @dividend:  64bit dividend
+ * @divisor:   64bit divisor
+ * @remainder:  64bit remainder
+ *
+ * This implementation is a comparable to algorithm used by div64_u64.
+ * But this operation, which includes math for calculating the remainder,
+ * is kept distinct to avoid slowing down the div64_u64 operation on 32bit
+ * systems.
+ */
+#ifndef div64_u64_rem
+u64 div64_u64_rem(u64 dividend, u64 divisor, u64 *remainder)
+{
+       u32 high = divisor >> 32;
+       u64 quot;
+
+       if (high == 0) {
+               u32 rem32;
+               quot = div_u64_rem(dividend, divisor, &rem32);
+               *remainder = rem32;
+       } else {
+               int n = 1 + fls(high);
+               quot = div_u64(dividend >> n, divisor >> n);
+
+               if (quot != 0)
+                       quot--;
+
+               *remainder = dividend - quot * divisor;
+               if (*remainder >= divisor) {
+                       quot++;
+                       *remainder -= divisor;
+               }
+       }
+
+       return quot;
+}
+EXPORT_SYMBOL(div64_u64_rem);
+#endif
+
+/**
+ * div64_u64 - unsigned 64bit divide with 64bit divisor
+ * @dividend:  64bit dividend
+ * @divisor:   64bit divisor
+ *
+ * This implementation is a modified version of the algorithm proposed
+ * by the book 'Hacker's Delight'.  The original source and full proof
+ * can be found here and is available for use without restriction.
+ *
+ * 'http://www.hackersdelight.org/hdcodetxt/divDouble.c.txt'
+ */
+#ifndef div64_u64
+u64 div64_u64(u64 dividend, u64 divisor)
+{
+       u32 high = divisor >> 32;
+       u64 quot;
+
+       if (high == 0) {
+               quot = div_u64(dividend, divisor);
+       } else {
+               int n = 1 + fls(high);
+               quot = div_u64(dividend >> n, divisor >> n);
+
+               if (quot != 0)
+                       quot--;
+               if ((dividend - quot * divisor) >= divisor)
+                       quot++;
+       }
+
+       return quot;
+}
+EXPORT_SYMBOL(div64_u64);
+#endif
+
+/**
+ * div64_s64 - signed 64bit divide with 64bit divisor
+ * @dividend:  64bit dividend
+ * @divisor:   64bit divisor
+ */
+#ifndef div64_s64
+s64 div64_s64(s64 dividend, s64 divisor)
+{
+       s64 quot, t;
+
+       quot = div64_u64(abs(dividend), abs(divisor));
+       t = (dividend ^ divisor) >> 63;
+
+       return (quot ^ t) - t;
+}
+EXPORT_SYMBOL(div64_s64);
+#endif
+
+#endif /* BITS_PER_LONG == 32 */
+
+/*
+ * Iterative div/mod for use when dividend is not expected to be much
+ * bigger than divisor.
+ */
+u32 iter_div_u64_rem(u64 dividend, u32 divisor, u64 *remainder)
+{
+       return __iter_div_u64_rem(dividend, divisor, remainder);
+}
+EXPORT_SYMBOL(iter_div_u64_rem);