-#if 0 /* old code... */
-static inline int fls(unsigned int a)
-{
- int r;
-
- asm volatile ("bfffo %1{#0,#32},%0"
- : "=d" (r) : "md" (a));
- return r;
-}
-
-/* fls = "find last set" (cf. ffs(3)) */
-static inline int fls128(const int128 a)
-{
- if (a[MSW128])
- return fls(a[MSW128]);
- if (a[NMSW128])
- return fls(a[NMSW128]) + 32;
- /* XXX: it probably never gets beyond this point in actual
- use, but that's indicative of a more general problem in the
- algorithm (i.e. as per the actual 68881 implementation, we
- really only need at most 67 bits of precision [plus
- overflow]) so I'm not going to fix it. */
- if (a[NLSW128])
- return fls(a[NLSW128]) + 64;
- if (a[LSW128])
- return fls(a[LSW128]) + 96;
- else
- return -1;
-}
-
-static inline int zerop128(const int128 a)
-{
- return !(a[LSW128] | a[NLSW128] | a[NMSW128] | a[MSW128]);
-}
-
-static inline int nonzerop128(const int128 a)
-{
- return (a[LSW128] | a[NLSW128] | a[NMSW128] | a[MSW128]);
-}
-
-/* Addition and subtraction */
-/* Do these in "pure" assembly, because "extended" asm is unmanageable
- here */
-static inline void add128(const int128 a, int128 b)
-{
- /* rotating carry flags */
- unsigned int carry[2];
-
- carry[0] = a[LSW128] > (0xffffffff - b[LSW128]);
- b[LSW128] += a[LSW128];
-
- carry[1] = a[NLSW128] > (0xffffffff - b[NLSW128] - carry[0]);
- b[NLSW128] = a[NLSW128] + b[NLSW128] + carry[0];
-
- carry[0] = a[NMSW128] > (0xffffffff - b[NMSW128] - carry[1]);
- b[NMSW128] = a[NMSW128] + b[NMSW128] + carry[1];
-
- b[MSW128] = a[MSW128] + b[MSW128] + carry[0];
-}
-
-/* Note: assembler semantics: "b -= a" */
-static inline void sub128(const int128 a, int128 b)
-{
- /* rotating borrow flags */
- unsigned int borrow[2];
-
- borrow[0] = b[LSW128] < a[LSW128];
- b[LSW128] -= a[LSW128];
-
- borrow[1] = b[NLSW128] < a[NLSW128] + borrow[0];
- b[NLSW128] = b[NLSW128] - a[NLSW128] - borrow[0];
-
- borrow[0] = b[NMSW128] < a[NMSW128] + borrow[1];
- b[NMSW128] = b[NMSW128] - a[NMSW128] - borrow[1];
-
- b[MSW128] = b[MSW128] - a[MSW128] - borrow[0];
-}
-
-/* Poor man's 64-bit expanding multiply */
-static inline void mul64(unsigned long long a, unsigned long long b, int128 c)
-{
- unsigned long long acc;
- int128 acc128;
-
- zero128(acc128);
- zero128(c);
-
- /* first the low words */
- if (LO_WORD(a) && LO_WORD(b)) {
- acc = (long long) LO_WORD(a) * LO_WORD(b);
- c[NLSW128] = HI_WORD(acc);
- c[LSW128] = LO_WORD(acc);
- }
- /* Next the high words */
- if (HI_WORD(a) && HI_WORD(b)) {
- acc = (long long) HI_WORD(a) * HI_WORD(b);
- c[MSW128] = HI_WORD(acc);
- c[NMSW128] = LO_WORD(acc);
- }
- /* The middle words */
- if (LO_WORD(a) && HI_WORD(b)) {
- acc = (long long) LO_WORD(a) * HI_WORD(b);
- acc128[NMSW128] = HI_WORD(acc);
- acc128[NLSW128] = LO_WORD(acc);
- add128(acc128, c);
- }
- /* The first and last words */
- if (HI_WORD(a) && LO_WORD(b)) {
- acc = (long long) HI_WORD(a) * LO_WORD(b);
- acc128[NMSW128] = HI_WORD(acc);
- acc128[NLSW128] = LO_WORD(acc);
- add128(acc128, c);
- }
-}
-
-/* Note: unsigned */
-static inline int cmp128(int128 a, int128 b)
-{
- if (a[MSW128] < b[MSW128])
- return -1;
- if (a[MSW128] > b[MSW128])
- return 1;
- if (a[NMSW128] < b[NMSW128])
- return -1;
- if (a[NMSW128] > b[NMSW128])
- return 1;
- if (a[NLSW128] < b[NLSW128])
- return -1;
- if (a[NLSW128] > b[NLSW128])
- return 1;
-
- return (signed) a[LSW128] - b[LSW128];
-}
-
-inline void div128(int128 a, int128 b, int128 c)
-{
- int128 mask;
-
- /* Algorithm:
-
- Shift the divisor until it's at least as big as the
- dividend, keeping track of the position to which we've
- shifted it, i.e. the power of 2 which we've multiplied it
- by.
-
- Then, for this power of 2 (the mask), and every one smaller
- than it, subtract the mask from the dividend and add it to
- the quotient until the dividend is smaller than the raised
- divisor. At this point, divide the dividend and the mask
- by 2 (i.e. shift one place to the right). Lather, rinse,
- and repeat, until there are no more powers of 2 left. */
-
- /* FIXME: needless to say, there's room for improvement here too. */
-
- /* Shift up */
- /* XXX: since it just has to be "at least as big", we can
- probably eliminate this horribly wasteful loop. I will
- have to prove this first, though */
- set128(0, 0, 0, 1, mask);
- while (cmp128(b, a) < 0 && !btsthi128(b)) {
- lslone128(b);
- lslone128(mask);
- }
-
- /* Shift down */
- zero128(c);
- do {
- if (cmp128(a, b) >= 0) {
- sub128(b, a);
- add128(mask, c);
- }
- lsrone128(mask);
- lsrone128(b);
- } while (nonzerop128(mask));
-
- /* The remainder is in a... */
-}
-#endif
-