Nailify.

11 files changed, 115 insertions, 50 deletions

diff --git a/mpf/cmp_si.c b/mpf/cmp_si.c
index d0a1c4c63..6266b953a 100644
--- a/mpf/cmp_si.c
+++ b/mpf/cmp_si.c
@@ -23,24 +23,25 @@ MA 02111-1307, USA. */
 #include "gmp-impl.h"
 
 int
-mpf_cmp_si (mpf_srcptr u, long int vslimb)
+mpf_cmp_si (mpf_srcptr u, long int vval)
 {
   mp_srcptr up;
   mp_size_t usize;
   mp_exp_t uexp;
+  mp_limb_t ulimb;
   int usign;
 
   uexp = u->_mp_exp;
   usize = u->_mp_size;
 
   /* 1. Are the signs different?  */
-  if ((usize < 0) == (vslimb < 0)) /* don't use xor, type size may differ */
+  if ((usize < 0) == (vval < 0)) /* don't use xor, type size may differ */
     {
       /* U and V are both non-negative or both negative.  */
       if (usize == 0)
-	/* vslimb >= 0 */
-	return -(vslimb != 0);
-      if (vslimb == 0)
+	/* vval >= 0 */
+	return -(vval != 0);
+      if (vval == 0)
 	/* usize >= 0 */
 	return usize != 0;
       /* Fall out.  */
@@ -54,37 +55,52 @@ mpf_cmp_si (mpf_srcptr u, long int vslimb)
   /* U and V have the same sign and are both non-zero.  */
 
   usign = usize >= 0 ? 1 : -1;
+  usize = ABS (usize);
+  vval = ABS (vval);
 
   /* 2. Are the exponents different (V's exponent == 1)?  */
+#if GMP_NAIL_BITS != 0
+  if (uexp > 1 + (vval > GMP_NUMB_MAX))
+    return usign;
+  if (uexp < 1 + (vval > GMP_NUMB_MAX))
+    return -usign;
+#else
   if (uexp > 1)
     return usign;
   if (uexp < 1)
     return -usign;
-
-  usize = ABS (usize);
-  vslimb = ABS (vslimb);
+#endif
 
   up = u->_mp_d;
 
+  ulimb = up[usize - 1];
+#if GMP_NAIL_BITS != 0
+  if (usize >= 2 && uexp == 2)
+    {
+      if ((ulimb >> GMP_NAIL_BITS) != 0)
+	return 1;
+      ulimb = (ulimb << GMP_NUMB_BITS) | up[usize - 2];
+      usize--;
+    }
+#endif
+  usize--;
+
   /* 3. Compare the most significant mantissa limb with V.  */
-  if (up[usize - 1] > (unsigned long) vslimb)
+  if (ulimb > (unsigned long) vval)
     return usign;
-  else if (up[usize - 1] < (unsigned long) vslimb)
+  else if (ulimb < (unsigned long) vval)
     return -usign;
 
-#define STRICT_MPF_NORMALIZATION 0
-#if ! STRICT_MPF_NORMALIZATION
   /* Ignore zeroes at the low end of U.  */
   while (*up == 0)
     {
       up++;
       usize--;
     }
-#endif
 
   /* 4. Now, if the number of limbs are different, we have a difference
      since we have made sure the trailing limbs are not zero.  */
-  if (usize > 1)
+  if (usize > 0)
     return usign;
 
   /* Wow, we got zero even if we tried hard to avoid it.  */
diff --git a/mpf/cmp_ui.c b/mpf/cmp_ui.c
index cbb17bbac..47ca597a4 100644
--- a/mpf/cmp_ui.c
+++ b/mpf/cmp_ui.c
@@ -23,11 +23,12 @@ MA 02111-1307, USA. */
 #include "gmp-impl.h"
 
 int
-mpf_cmp_ui (mpf_srcptr u, unsigned long int vlimb)
+mpf_cmp_ui (mpf_srcptr u, unsigned long int vval)
 {
   mp_srcptr up;
   mp_size_t usize;
   mp_exp_t uexp;
+  mp_limb_t ulimb;
 
   uexp = u->_mp_exp;
   usize = u->_mp_size;
@@ -37,36 +38,52 @@ mpf_cmp_ui (mpf_srcptr u, unsigned long int vlimb)
     return -1;
   /* We rely on usize being non-negative in the code that follows.  */
 
-  if (vlimb == 0)
+  if (vval == 0)
     return usize != 0;
 
   /* 2. Are the exponents different (V's exponent == 1)?  */
+#if GMP_NAIL_BITS != 0
+  if (uexp > 1 + (vval > GMP_NUMB_MAX))
+    return 1;
+  if (uexp < 1 + (vval > GMP_NUMB_MAX))
+    return -1;
+#else
   if (uexp > 1)
     return 1;
   if (uexp < 1)
     return -1;
+#endif
 
   up = u->_mp_d;
 
+  ulimb = up[usize - 1];
+#if GMP_NAIL_BITS != 0
+  if (usize >= 2 && uexp == 2)
+    {
+      if ((ulimb >> GMP_NAIL_BITS) != 0)
+	return 1;
+      ulimb = (ulimb << GMP_NUMB_BITS) | up[usize - 2];
+      usize--;
+    }
+#endif
+  usize--;
+
   /* 3. Compare the most significant mantissa limb with V.  */
-  if (up[usize - 1] > vlimb)
+  if (ulimb > vval)
     return 1;
-  else if (up[usize - 1] < vlimb)
+  else if (ulimb < vval)
     return -1;
 
-#define STRICT_MPF_NORMALIZATION 0
-#if ! STRICT_MPF_NORMALIZATION
   /* Ignore zeroes at the low end of U.  */
   while (*up == 0)
     {
       up++;
       usize--;
     }
-#endif
 
   /* 4. Now, if the number of limbs are different, we have a difference
      since we have made sure the trailing limbs are not zero.  */
-  if (usize > 1)
+  if (usize > 0)
     return 1;
 
   /* Wow, we got zero even if we tried hard to avoid it.  */
diff --git a/mpf/div.c b/mpf/div.c
index c8ce73442..0f298a554 100644
--- a/mpf/div.c
+++ b/mpf/div.c
@@ -83,13 +83,14 @@ mpf_div (mpf_ptr r, mpf_srcptr u, mpf_srcptr v)
 
 
   /* Normalize the divisor and the dividend.  */
-  if (! (vp[vsize-1] & MP_LIMB_T_HIGHBIT))
+  if ((vp[vsize-1] & GMP_NUMB_HIGHBIT) == 0)
     {
       mp_ptr tmp;
       mp_limb_t nlimb;
       unsigned normalization_steps;
 
       count_leading_zeros (normalization_steps, vp[vsize - 1]);
+      normalization_steps -= GMP_NAIL_BITS;
 
       /* Shift up the divisor setting the most significant bit of
 	 the most significant limb.  Use temporary storage not to clobber
diff --git a/mpf/div_2exp.c b/mpf/div_2exp.c
index 870abefe3..6f70aedd5 100644
--- a/mpf/div_2exp.c
+++ b/mpf/div_2exp.c
@@ -44,7 +44,7 @@ mpf_div_2exp (mpf_ptr r, mpf_srcptr u, unsigned long int exp)
   abs_usize = ABS (usize);
   up = u->_mp_d;
 
-  if (exp % BITS_PER_MP_LIMB == 0)
+  if (exp % GMP_NUMB_BITS == 0)
     {
       prec++;			/* retain more precision here as we don't need
 				   to account for carry-out here */
@@ -55,7 +55,7 @@ mpf_div_2exp (mpf_ptr r, mpf_srcptr u, unsigned long int exp)
 	}
       if (rp != up)
 	MPN_COPY_INCR (rp, up, abs_usize);
-      r->_mp_exp = uexp - exp / BITS_PER_MP_LIMB;
+      r->_mp_exp = uexp - exp / GMP_NUMB_BITS;
     }
   else
     {
@@ -68,19 +68,20 @@ mpf_div_2exp (mpf_ptr r, mpf_srcptr u, unsigned long int exp)
 	  /* Use mpn_rshift since mpn_lshift operates downwards, and we
 	     therefore would clobber part of U before using that part, in case
 	     R is the same variable as U.  */
-	  cy_limb = mpn_rshift (rp + 1, up, abs_usize, exp % BITS_PER_MP_LIMB);
+	  cy_limb = mpn_rshift (rp + 1, up, abs_usize, exp % GMP_NUMB_BITS);
 	  rp[0] = cy_limb;
 	  adj = rp[abs_usize] != 0;
 	}
       else
 	{
-	  cy_limb = mpn_lshift (rp, up, abs_usize, (-exp) % BITS_PER_MP_LIMB);
+	  cy_limb = mpn_lshift (rp, up, abs_usize,
+				GMP_NUMB_BITS - exp % GMP_NUMB_BITS);
 	  rp[abs_usize] = cy_limb;
 	  adj = cy_limb != 0;
 	}
 
       abs_usize += adj;
-      r->_mp_exp = uexp - exp / BITS_PER_MP_LIMB - 1 + adj;
+      r->_mp_exp = uexp - exp / GMP_NUMB_BITS - 1 + adj;
     }
   r->_mp_size = usize >= 0 ? abs_usize : -abs_usize;
 }
diff --git a/mpf/div_ui.c b/mpf/div_ui.c
index 74405afd7..8aed3dbb7 100644
--- a/mpf/div_ui.c
+++ b/mpf/div_ui.c
@@ -36,6 +36,21 @@ mpf_div_ui (mpf_ptr r, mpf_srcptr u, unsigned long int v)
   mp_exp_t rexp;
   TMP_DECL (marker);
 
+#if GMP_NAIL_BITS != 0
+  if (v > GMP_NUMB_MAX)
+    {
+      mpf_t vf;
+      mp_limb_t vl[2];
+      SIZ(vf) = 2;
+      EXP(vf) = 2;
+      PTR(vf) = vl;
+      vl[0] = v & GMP_NUMB_MASK;
+      vl[1] = v >> GMP_NUMB_BITS;
+      mpf_div (r, u, vf);
+      return;
+    }
+#endif
+
   usize = u->_mp_size;
   sign_quotient = usize;
   usize = ABS (usize);
diff --git a/mpf/eq.c b/mpf/eq.c
index 52dd366ef..e33f51351 100644
--- a/mpf/eq.c
+++ b/mpf/eq.c
@@ -83,13 +83,13 @@ mpf_eq (mpf_srcptr u, mpf_srcptr v, unsigned long int n_bits)
 
   if (usize > vsize)
     {
-      if (vsize * BITS_PER_MP_LIMB < n_bits)
+      if (vsize * GMP_NUMB_BITS < n_bits)
 	return 0;		/* surely too different */
       size = vsize;
     }
   else if (vsize > usize)
     {
-      if (usize * BITS_PER_MP_LIMB < n_bits)
+      if (usize * GMP_NUMB_BITS < n_bits)
 	return 0;		/* surely too different */
       size = usize;
     }
@@ -98,8 +98,8 @@ mpf_eq (mpf_srcptr u, mpf_srcptr v, unsigned long int n_bits)
       size = usize;
     }
 
-  if (size > (n_bits + BITS_PER_MP_LIMB - 1) / BITS_PER_MP_LIMB)
-    size = (n_bits + BITS_PER_MP_LIMB - 1) / BITS_PER_MP_LIMB;
+  if (size > (n_bits + GMP_NUMB_BITS - 1) / GMP_NUMB_BITS)
+    size = (n_bits + GMP_NUMB_BITS - 1) / GMP_NUMB_BITS;
 
   up += usize - size;
   vp += vsize - size;
diff --git a/mpf/get_d.c b/mpf/get_d.c
index 42220d54a..d2a56e2af 100644
--- a/mpf/get_d.c
+++ b/mpf/get_d.c
@@ -43,7 +43,7 @@ mpf_get_d (mpf_srcptr src)
   for (i = 2; i <= n_limbs_to_use; i++)
     res = res * MP_BASE_AS_DOUBLE + qp[size - i];
 
-  res = __gmp_scale2 (res, (EXP(src) - n_limbs_to_use) * BITS_PER_MP_LIMB);
+  res = __gmp_scale2 (res, (EXP(src) - n_limbs_to_use) * GMP_NUMB_BITS);
 
   return negative ? -res : res;
 }
diff --git a/mpf/get_d_2exp.c b/mpf/get_d_2exp.c
index 92e3a2edc..cc5aced62 100644
--- a/mpf/get_d_2exp.c
+++ b/mpf/get_d_2exp.c
@@ -49,7 +49,7 @@ mpf_get_d_2exp (signed long int *exp2, mpf_srcptr src)
   for (i = 1; i < n_limbs_to_use; i++)
     res = (res + qp[i]) / MP_BASE_AS_DOUBLE;
   count_leading_zeros (cnt, qp[n_limbs_to_use - 1]);
-  *exp2 = EXP(src) * BITS_PER_MP_LIMB - cnt;
+  *exp2 = EXP(src) * GMP_NUMB_BITS - cnt + GMP_NAIL_BITS;
   res = res * ((mp_limb_t) 1 << cnt);
 
   return negative ? -res : res;
diff --git a/mpf/get_si.c b/mpf/get_si.c
index 9246685c3..006966df0 100644
--- a/mpf/get_si.c
+++ b/mpf/get_si.c
@@ -35,9 +35,10 @@ MA 02111-1307, USA.
 long
 mpf_get_si (mpf_srcptr f)
 {
-  mp_exp_t  exp;
-  int       size, abs_size;
-  long      n;
+  mp_exp_t exp;
+  mp_size_t size, abs_size;
+  mp_ptr fp;
+  mp_limb_t fl;
 
   size = SIZ (f);
   if (size == 0)
@@ -53,10 +54,17 @@ mpf_get_si (mpf_srcptr f)
   if (exp > abs_size)
     return 0L;
 
-  n = (long) PTR(f)[abs_size-exp];
+  fp = PTR(f);
+  fl = fp[abs_size - exp];
+
+#if GMP_NAIL_BITS != 0
+  if (ULONG_MAX > GMP_NUMB_MAX != 0 && exp > 1)
+    fl |= fp[abs_size - exp + 1] << GMP_NUMB_BITS;
+#endif
+
   if (size > 0)
-    return n & LONG_MAX;
+    return fl & LONG_MAX;
   else
     /* this form necessary to correctly handle -0x80..00 */
-    return ~ ((n - 1) & LONG_MAX);
+    return ~ ((fl - 1) & LONG_MAX);
 }
diff --git a/mpf/get_str.c b/mpf/get_str.c
index c57fc0f5c..b61e230ef 100644
--- a/mpf/get_str.c
+++ b/mpf/get_str.c
@@ -92,7 +92,7 @@ mpf_get_str (char *digit_ptr, mp_exp_t *exp, int base, size_t n_digits, mpf_srcp
 	 desired.  We could probably decrease both this, and avoid the +1
 	 for setting prec above.  */
       prec = 2 + (mp_size_t)
-	(n_digits / (BITS_PER_MP_LIMB * __mp_bases[base].chars_per_bit_exactly));
+	(n_digits / (GMP_NUMB_BITS * __mp_bases[base].chars_per_bit_exactly));
 #endif
   }
 
@@ -136,7 +136,7 @@ mpf_get_str (char *digit_ptr, mp_exp_t *exp, int base, size_t n_digits, mpf_srcp
 #endif
 
       count_leading_zeros (cnt, up[usize - 1]);
-      exp_in_base = (((double) uexp * BITS_PER_MP_LIMB - cnt)
+      exp_in_base = (((double) uexp * GMP_NUMB_BITS - cnt + GMP_NAIL_BITS)
 		     * __mp_bases[base].chars_per_bit_exactly);
       exp_in_base += 1;
 
@@ -147,7 +147,7 @@ mpf_get_str (char *digit_ptr, mp_exp_t *exp, int base, size_t n_digits, mpf_srcp
       rp[0] = base;
       rsize = 1;
       count_leading_zeros (cnt, exp_in_base);
-      for (i = BITS_PER_MP_LIMB - cnt - 2; i >= 0; i--)
+      for (i = GMP_LIMB_BITS - cnt - 2; i >= 0; i--)
 	{
 	  mpn_sqr_n (tp, rp, rsize);
 	  rsize = 2 * rsize;
@@ -171,6 +171,7 @@ mpf_get_str (char *digit_ptr, mp_exp_t *exp, int base, size_t n_digits, mpf_srcp
 	}
 
       count_leading_zeros (cnt, rp[rsize - 1]);
+      cnt -= GMP_NAIL_BITS;
       if (cnt != 0)
 	{
 	  mpn_lshift (rp, rp, rsize, cnt);
@@ -239,7 +240,7 @@ mpf_get_str (char *digit_ptr, mp_exp_t *exp, int base, size_t n_digits, mpf_srcp
 
       uexp = -uexp;
       count_leading_zeros (cnt, up[usize - 1]);
-      exp_in_base = (((double) uexp * BITS_PER_MP_LIMB + cnt - 1)
+      exp_in_base = (((double) uexp * GMP_NUMB_BITS + cnt - GMP_NAIL_BITS - 1)
 		     * __mp_bases[base].chars_per_bit_exactly);
       if (exp_in_base < 0)
 	exp_in_base = 0;
@@ -253,7 +254,7 @@ mpf_get_str (char *digit_ptr, mp_exp_t *exp, int base, size_t n_digits, mpf_srcp
 	  rp[0] = base;
 	  rsize = 1;
 	  count_leading_zeros (cnt, exp_in_base);
-	  for (i = BITS_PER_MP_LIMB - cnt - 2; i >= 0; i--)
+	  for (i = GMP_LIMB_BITS - cnt - 2; i >= 0; i--)
 	    {
 	      mpn_sqr_n (tp, rp, rsize);
 	      rsize = 2 * rsize;
@@ -305,7 +306,7 @@ mpf_get_str (char *digit_ptr, mp_exp_t *exp, int base, size_t n_digits, mpf_srcp
   if (big_base < 10)		/* logarithm of base when power of two */
     {
       int logbase = big_base;
-      if (dig_per_u * logbase == BITS_PER_MP_LIMB)
+      if (dig_per_u * logbase == GMP_NUMB_BITS)
 	dig_per_u--;
       big_base = (mp_limb_t) 1 << (dig_per_u * logbase);
       /* fall out to general code... */
@@ -318,9 +319,9 @@ mpf_get_str (char *digit_ptr, mp_exp_t *exp, int base, size_t n_digits, mpf_srcp
 
   /* Allocate temporary digit space.  We can't put digits directly in the user
      area, since we generate more digits than requested.  (We allocate
-     BITS_PER_MP_LIMB extra bytes because of the digit block nature of the
+     GMP_LIMB_BITS extra bytes because of the digit block nature of the
      conversion.)  */
-  tstr = (unsigned char *) TMP_ALLOC (n_digits + BITS_PER_MP_LIMB + 3);
+  tstr = (unsigned char *) TMP_ALLOC (n_digits + GMP_LIMB_BITS + 3);
 
   for (digits_computed_so_far = 0; digits_computed_so_far < n_digits + 3;
        digits_computed_so_far += dig_per_u)
@@ -337,6 +338,12 @@ mpf_get_str (char *digit_ptr, mp_exp_t *exp, int base, size_t n_digits, mpf_srcp
 
       cy = mpn_mul_1 (rp, rp, rsize, big_base);
 
+      if (! POW2_P (GMP_NUMB_BITS))
+	{
+	  if (digits_computed_so_far == 0 && cy == 0)
+	    cy = mpn_mul_1 (rp, rp, rsize, big_base);
+	}
+
       ASSERT_ALWAYS (! (digits_computed_so_far == 0 && cy == 0));
 
       /* Convert N1 from BIG_BASE to a string of digits in BASE
diff --git a/mpf/get_ui.c b/mpf/get_ui.c
index df65a1f6c..c4a04706e 100644
--- a/mpf/get_ui.c
+++ b/mpf/get_ui.c
@@ -37,8 +37,8 @@ MA 02111-1307, USA.
 unsigned long
 mpf_get_ui (mpf_srcptr f)
 {
-  int        size, abs_size;
-  mp_exp_t   exp;
+  mp_size_t size, abs_size;
+  mp_exp_t exp;
 
   size = SIZ (f);
   if (size == 0)