5 * If CONFIG_M68000=y (original mc68000/010), this file is #included
6 * to work around the lack of a MULU.L instruction.
9 #define HAVE_ARCH__HASH_32 1
11 * While it would be legal to substitute a different hash operation
12 * entirely, let's keep it simple and just use an optimized multiply
13 * by GOLDEN_RATIO_32 = 0x61C88647.
15 * The best way to do that appears to be to multiply by 0x8647 with
16 * shifts and adds, and use mulu.w to multiply the high half by 0x61C8.
18 * Because the 68000 has multi-cycle shifts, this addition chain is
19 * chosen to minimise the shift distances.
21 * Despite every attempt to spoon-feed it simple operations, GCC
22 * 6.1.1 doggedly insists on doing annoying things like converting
23 * "lsl.l #2,<reg>" (12 cycles) to two adds (8+8 cycles).
25 * It also likes to notice two shifts in a row, like "a = x << 2" and
26 * "a <<= 7", and convert that to "a = x << 9". But shifts longer
27 * than 8 bits are extra-slow on m68k, so that's a lose.
29 * Since the 68000 is a very simple in-order processor with no
30 * instruction scheduling effects on execution time, we can safely
31 * take it out of GCC's hands and write one big asm() block.
33 * Without calling overhead, this operation is 30 bytes (14 instructions
34 * plus one immediate constant) and 166 cycles.
36 * (Because %2 is fetched twice, it can't be postincrement, and thus it
37 * can't be a fully general "g" or "m". Register is preferred, but
38 * offsettable memory or immediate will work.)
40 static inline u32 __attribute_const__ __hash_32(u32 x)
44 asm( "move.l %2,%0" /* a = x * 0x0001 */
45 "\n lsl.l #2,%0" /* a = x * 0x0004 */
47 "\n lsl.l #7,%0" /* a = x * 0x0200 */
48 "\n add.l %2,%0" /* a = x * 0x0201 */
49 "\n add.l %0,%1" /* b = x * 0x0205 */
50 "\n add.l %0,%0" /* a = x * 0x0402 */
51 "\n add.l %0,%1" /* b = x * 0x0607 */
52 "\n lsl.l #5,%0" /* a = x * 0x8040 */
53 : "=&d,d" (a), "=&r,r" (b)
54 : "r,roi?" (x)); /* a+b = x*0x8647 */
56 return ((u16)(x*0x61c8) << 16) + a + b;
59 #endif /* _ASM_HASH_H */