7 * Copyright (C) 1994 Hamish Macdonald
8 * Copyright (C) 2004 Greg Ungerer <gerg@uclinux.com>
10 * Delay routines, using a pre-computed "loops_per_jiffy" value.
13 #if defined(CONFIG_COLDFIRE)
15 * The ColdFire runs the delay loop at significantly different speeds
16 * depending upon long word alignment or not. We'll pad it to
17 * long word alignment which is the faster version.
18 * The 0x4a8e is of course a 'tstl %fp' instruction. This is better
19 * than using a NOP (0x4e71) instruction because it executes in one
20 * cycle not three and doesn't allow for an arbitrary delay waiting
21 * for bus cycles to finish. Also fp/a6 isn't likely to cause a
22 * stall waiting for the register to become valid if such is added
23 * to the coldfire at some stage.
25 #define DELAY_ALIGN ".balignw 4, 0x4a8e\n\t"
28 * No instruction alignment required for other m68k types.
33 static inline void __delay(unsigned long loops)
35 __asm__ __volatile__ (
43 extern void __bad_udelay(void);
46 #ifdef CONFIG_CPU_HAS_NO_MULDIV64
48 * The simpler m68k and ColdFire processors do not have a 32*32->64
49 * multiply instruction. So we need to handle them a little differently.
50 * We use a bit of shifting and a single 32*32->32 multiply to get close.
51 * This is a macro so that the const version can factor out the first
54 #define HZSCALE (268435456 / (1000000 / HZ))
56 #define __const_udelay(u) \
57 __delay(((((u) * HZSCALE) >> 11) * (loops_per_jiffy >> 11)) >> 6)
61 static inline void __xdelay(unsigned long xloops)
65 __asm__ ("mulul %2,%0:%1"
66 : "=d" (xloops), "=d" (tmp)
67 : "d" (xloops), "1" (loops_per_jiffy));
72 * The definition of __const_udelay is specifically made a macro so that
73 * the const factor (4295 = 2**32 / 1000000) can be optimized out when
74 * the delay is a const.
76 #define __const_udelay(n) (__xdelay((n) * 4295))
80 static inline void __udelay(unsigned long usecs)
82 __const_udelay(usecs);
86 * Use only for very small delays ( < 1 msec). Should probably use a
87 * lookup table, really, as the multiplications take much too long with
88 * short delays. This is a "reasonable" implementation, though (and the
89 * first constant multiplications gets optimized away if the delay is
92 #define udelay(n) (__builtin_constant_p(n) ? \
93 ((n) > 20000 ? __bad_udelay() : __const_udelay(n)) : __udelay(n))
98 * ((((HZSCALE) >> 11) * (loops_per_jiffy >> 11)) >> 6) is the number of loops
101 * 1000 / ((((HZSCALE) >> 11) * (loops_per_jiffy >> 11)) >> 6) is the number of
102 * nanoseconds per loop
104 * So n / ( 1000 / ((((HZSCALE) >> 11) * (loops_per_jiffy >> 11)) >> 6) ) would
105 * be the number of loops for n nanoseconds
109 * The simpler m68k and ColdFire processors do not have a 32*32->64
110 * multiply instruction. So we need to handle them a little differently.
111 * We use a bit of shifting and a single 32*32->32 multiply to get close.
112 * This is a macro so that the const version can factor out the first
113 * multiply and shift.
115 #define HZSCALE (268435456 / (1000000 / HZ))
117 #define ndelay(n) __delay(DIV_ROUND_UP((n) * ((((HZSCALE) >> 11) * (loops_per_jiffy >> 11)) >> 6), 1000));
119 #endif /* defined(_M68K_DELAY_H) */