Change sub-quadratic divexact task to want this for mpn_bdivmod which

will then help divexact, gcd and hopefully redc.
Add mpn_gcd_1 to use exact division reduction not mpn_mod_1.
Add --enable-abi idea.

1 files changed, 19 insertions, 3 deletions

diff --git a/doc/tasks.html b/doc/tasks.html
index f914c2413..f574be62d 100644
--- a/doc/tasks.html
+++ b/doc/tasks.html
@@ -143,10 +143,19 @@
      targets with native versions of those routines.  Need to be careful not to
      pass size==0.  Is some code sharing possible between the <code>mpz</code>
      routines?
-<li> <code>mpz_divexact</code> should use a divide and conquer like the normal
+<li> <code>mpn_bdivmod</code> should use a divide and conquer like the normal
      division.  See "Exact Division with Karatsuba Complexity" by Jebelean for
-     a (brief) description.  Might be possible to share code with a divide and
-     conquer REDC.
+     a (brief) description.  This will benefit <code>mpz_divexact</code>
+     immediately, and <code>mpn_gcd</code> on large unequal size operands.
+     REDC should be able to use it too.
+<li> <code>mpn_gcd_1</code>, <code>mpz_kronecker_ui</code> etc should be able
+     to do an exact division style reduction initially, rather than
+     <code>mpn_mod_1</code>.  This would be the same as <code>mpn_gcd</code>
+     does using <code>mpn_bdivmod</code>.  This is still two multiplies per
+     limb, but is simpler and should be faster than mul by inverse division.
+     Perhaps a function <code>mpn_modexact_1</code>, being the remainder part
+     of an <code>mpn_divexact_1</code>.  Creating the modular inverse will take
+     a few cycles, so it might be only for say 4 limbs and up.
 </ul>
 
 
@@ -306,6 +315,13 @@
      be conditionalized, and the appropriate form could be selected with the
      <code>HAVE_NATIVE</code> scheme if/when the code switches to use a
      <code>PROLOGUE</code> style.
+<li> Think about a --enable-abi=FOO to select which abi to use on systems that
+     support multiple calling conventions and/or data sizes.  The default could
+     be "guess", other choices could be made by users wanting compatibility
+     with other object code.  Doing the right thing with where and how to
+     install library files probably needs a lot of help from automake and
+     libtool.  A few simple cases like /usr/lib/sparcv9 might be easy enough
+     though.
 </ul>
 
 <p> In general, getting the exact right configuration, passing the