[project @ 1996-01-08 20:28:12 by partain]
[ghc-hetmet.git] / ghc / runtime / gmp / mpn_dm_1.c
diff --git a/ghc/runtime/gmp/mpn_dm_1.c b/ghc/runtime/gmp/mpn_dm_1.c
new file mode 100644 (file)
index 0000000..af39124
--- /dev/null
@@ -0,0 +1,185 @@
+/* mpn_divmod_1(quot_ptr, dividend_ptr, dividend_size, divisor_limb) --
+   Divide (DIVIDEND_PTR,,DIVIDEND_SIZE) by DIVISOR_LIMB.
+   Write DIVIDEND_SIZE limbs of quotient at QUOT_PTR.
+   Return the single-limb remainder.
+   There are no constraints on the value of the divisor.
+
+   QUOT_PTR and DIVIDEND_PTR might point to the same limb.
+
+Copyright (C) 1991, 1993 Free Software Foundation, Inc.
+
+This file is part of the GNU MP Library.
+
+The GNU MP Library is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2, or (at your option)
+any later version.
+
+The GNU MP Library is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the GNU MP Library; see the file COPYING.  If not, write to
+the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.  */
+
+#include "gmp.h"
+#include "gmp-impl.h"
+#include "longlong.h"
+
+#ifndef UMUL_TIME
+#define UMUL_TIME 1
+#endif
+
+#ifndef UDIV_TIME
+#define UDIV_TIME UMUL_TIME
+#endif
+
+#if UDIV_TIME > 2 * UMUL_TIME
+#undef UDIV_NEEDS_NORMALIZATION
+#define UDIV_NEEDS_NORMALIZATION 1
+#endif
+
+#define udiv_qrnnd_preinv(q, r, nh, nl, d, di) \
+  do {                                                                 \
+    unsigned long int _q, _ql, _r;                                     \
+    unsigned long int _xh, _xl;                                                \
+    umul_ppmm (_q, _ql, (nh), (di));                                   \
+    _q += (nh);                        /* DI is 2**BITS_PER_MP_LIMB too small.  */\
+    umul_ppmm (_xh, _xl, _q, (d));                                     \
+    sub_ddmmss (_xh, _r, (nh), (nl), _xh, _xl);                                \
+    if (_xh != 0)                                                      \
+      {                                                                        \
+       sub_ddmmss (_xh, _r, _xh, _r, 0, (d));                          \
+       _q += 1;                                                        \
+       if (_xh != 0)                                                   \
+         {                                                             \
+           sub_ddmmss (_xh, _r, _xh, _r, 0, (d));                      \
+           _q += 1;                                                    \
+         }                                                             \
+      }                                                                        \
+    if (_r >= (d))                                                     \
+      {                                                                        \
+       _r -= (d);                                                      \
+       _q += 1;                                                        \
+      }                                                                        \
+    (r) = _r;                                                          \
+    (q) = _q;                                                          \
+  } while (0)
+
+mp_limb
+#ifdef __STDC__
+mpn_divmod_1 (mp_ptr quot_ptr,
+             mp_srcptr dividend_ptr, mp_size dividend_size,
+             unsigned long int divisor_limb)
+#else
+mpn_divmod_1 (quot_ptr, dividend_ptr, dividend_size, divisor_limb)
+     mp_ptr quot_ptr;
+     mp_srcptr dividend_ptr;
+     mp_size dividend_size;
+     unsigned long int divisor_limb;
+#endif
+{
+  mp_size i;
+  mp_limb n1, n0, r;
+
+  /* Botch: Should this be handled at all?  Rely on callers?  */
+  if (dividend_size == 0)
+    return 0;
+
+  if (UDIV_NEEDS_NORMALIZATION)
+    {
+      int normalization_steps;
+
+      count_leading_zeros (normalization_steps, divisor_limb);
+      if (normalization_steps != 0)
+       {
+         divisor_limb <<= normalization_steps;
+
+         n1 = dividend_ptr[dividend_size - 1];
+         r = n1 >> (BITS_PER_MP_LIMB - normalization_steps);
+
+         /* Possible optimization:
+            if (r == 0
+            && divisor_limb > ((n1 << normalization_steps)
+                            | (dividend_ptr[dividend_size - 2] >> ...)))
+            ...one division less...
+            [Don't forget to zero most sign. quotient limb!]  */
+
+         /* If multiplication is much faster than division, and the
+            dividend is large, pre-invert the divisor, and use
+            only multiplications in the inner loop.  */
+         if (UDIV_TIME > 2 * UMUL_TIME && dividend_size >= 4)
+           {
+             mp_limb divisor_limb_inverted;
+             int dummy;
+             /* Compute (2**64 - 2**32 * DIVISOR_LIMB) / DIVISOR_LIMB.
+                The result is an 33-bit approximation to 1/DIVISOR_LIMB,
+                with the most significant bit (weight 2**32) implicit.  */
+
+             /* Special case for DIVISOR_LIMB == 100...000.  */
+             if (divisor_limb << 1 == 0)
+               divisor_limb_inverted = ~0;
+             else
+               udiv_qrnnd (divisor_limb_inverted, dummy,
+                           -divisor_limb, 0, divisor_limb);
+
+             for (i = dividend_size - 2; i >= 0; i--)
+               {
+                 n0 = dividend_ptr[i];
+                 udiv_qrnnd_preinv (quot_ptr[i + 1], r, r,
+                                    ((n1 << normalization_steps)
+                                     | (n0 >> (BITS_PER_MP_LIMB - normalization_steps))),
+                                    divisor_limb, divisor_limb_inverted);
+                 n1 = n0;
+               }
+             udiv_qrnnd_preinv (quot_ptr[0], r, r,
+                         n1 << normalization_steps,
+                         divisor_limb, divisor_limb_inverted);
+             return r >> normalization_steps;
+           }
+         else
+           {
+             for (i = dividend_size - 2; i >= 0; i--)
+               {
+                 n0 = dividend_ptr[i];
+                 udiv_qrnnd (quot_ptr[i + 1], r, r,
+                             ((n1 << normalization_steps)
+                              | (n0 >> (BITS_PER_MP_LIMB - normalization_steps))),
+                             divisor_limb);
+                 n1 = n0;
+               }
+             udiv_qrnnd (quot_ptr[0], r, r,
+                         n1 << normalization_steps,
+                         divisor_limb);
+             return r >> normalization_steps;
+           }
+       }
+    }
+
+  /* No normalization needed, either because udiv_qrnnd doesn't require
+     it, or because DIVISOR_LIMB is already normalized.  */
+
+  i = dividend_size - 1;
+  r = dividend_ptr[i];
+
+  if (r >= divisor_limb)
+    {
+      r = 0;
+    }
+  else
+    {
+      /* Callers expect the quotient to be DIVIDEND_SIZE limbs.  Store
+        a leading zero to make that expectation come true.  */
+      quot_ptr[i] = 0;
+      i--;
+    }
+
+  for (; i >= 0; i--)
+    {
+      n0 = dividend_ptr[i];
+      udiv_qrnnd (quot_ptr[i], r, r, n0, divisor_limb);
+    }
+  return r;
+}