*
* Suppose %$x = x' + z 2^k$%, where %$k \ge n$%; then
* %$x \equiv x' + d z 2^{k-n} \pmod p$%. We can use this to trim the
- * representation of %$x$%; each time, we reduce %$x$% by a mutliple of
+ * representation of %$x$%; each time, we reduce %$x$% by a multiple of
* %$2^{k-n} p$%. We can do this in two passes: firstly by taking whole
* words off the top, and then (if necessary) by trimming the top word.
* Finally, if %$p \le x < 2^n$% then %$0 \le x - p < p$% and we're done.
* Arguments: @gfreduce *r@ = structure to fill in
* @mp *x@ = an integer
*
- * Returns: Zero if successful; nonzero on failure.
+ * Returns: Zero if successful; nonzero on failure. The current
+ * algorithm always succeeds when given positive @x@. Earlier
+ * versions used to fail on particular kinds of integers, but
+ * this is guaranteed not to happen any more.
*
* Use: Initializes a context structure for reduction.
*/
}
}
- /* --- This doesn't always work --- *
+ /* --- Fix up wrong-sided decompositions --- *
*
- * If %$d \ge 2^{n-1}$% then the above recurrence will output a subtraction
- * as the final instruction, which may sometimes underflow. (It interprets
- * such numbers as being in the form %$2^{n-1} + d$%.) This is clearly
- * bad, so detect the situation and fail gracefully.
+ * At this point, we haven't actually finished up the state machine
+ * properly. We stopped scanning just after bit %$n - 1$% -- the most
+ * significant one, which we know in advance must be set (since @x@ is
+ * strictly positive). Therefore we are either in state @X@ or @Z1@. In
+ * the former case, we have nothing to do. In the latter, there are two
+ * subcases to deal with. If there are no other instructions, then @x@ is
+ * a perfect power of two, and %$d = 0$%, so again there is nothing to do.
+ *
+ * In the remaining case, we have decomposed @x@ as %$2^{n-1} + d$%, for
+ * some positive %$d%, which is unfortunate: if we're asked to reduce
+ * %$2^n$%, say, we'll end up with %$-d$% (or would do, if we weren't
+ * sticking to unsigned arithmetic for good performance). So instead, we
+ * rewrite this as %$2^n - 2^{n-1} + d$% and everything will be good.
*/
- if (DA_LEN(&iv) && (DA(&iv)[DA_LEN(&iv) - 1].op & ~1u) == MPRI_SUB) {
- mp_drop(r->p);
- DA_DESTROY(&iv);
- return (-1);
+ if (st == Z1 && DA_LEN(&iv)) {
+ w = 1;
+ b = (bb + d)%MPW_BITS;
+ INSTR(MPRI_ADD | !!b, w, b);
}
#undef INSTR
/* --- @mpreduce_dump@ --- *
*
- * Arguments: @mpreduce *r@ = structure to dump
+ * Arguments: @const mpreduce *r@ = structure to dump
* @FILE *fp@ = file to dump on
*
* Returns: ---
* Use: Dumps a reduction context.
*/
-void mpreduce_dump(mpreduce *r, FILE *fp)
+void mpreduce_dump(const mpreduce *r, FILE *fp)
{
size_t i;
static const char *opname[] = { "add", "addshift", "sub", "subshift" };
/* --- @mpreduce_do@ --- *
*
- * Arguments: @mpreduce *r@ = reduction context
+ * Arguments: @const mpreduce *r@ = reduction context
* @mp *d@ = destination
* @mp *x@ = source
*
}
}
-mp *mpreduce_do(mpreduce *r, mp *d, mp *x)
+mp *mpreduce_do(const mpreduce *r, mp *d, mp *x)
{
mpw *v, *vl;
const mpreduce_instr *il;
/* --- @mpreduce_exp@ --- *
*
- * Arguments: @mpreduce *mr@ = pointer to reduction context
+ * Arguments: @const mpreduce *mr@ = pointer to reduction context
* @mp *d@ = fake destination
* @mp *a@ = base
* @mp *e@ = exponent
* Returns: Result, %$a^e \bmod m$%.
*/
-mp *mpreduce_exp(mpreduce *mr, mp *d, mp *a, mp *e)
+mp *mpreduce_exp(const mpreduce *mr, mp *d, mp *a, mp *e)
{
mp *x = MP_ONE;
mp *spare = (e->f & MP_BURN) ? MP_NEWSEC : MP_NEW;
/*----- Test rig ----------------------------------------------------------*/
-
#ifdef TEST_RIG
-#define MP(x) mp_readstring(MP_NEW, #x, 0, 0)
-
static int vreduce(dstr *v)
{
mp *d = *(mp **)v[0].buf;
~mdw / catacomb / blobdiff