[src/get_ld.c] Minor performance improvement.

git-svn-id: svn://scm.gforge.inria.fr/svn/mpfr/trunk@7311 280ebfd0-de03-0410-8827-d642c229c3f4

1 files changed, 6 insertions, 1 deletions

diff --git a/src/get_ld.c b/src/get_ld.c
index b6c05d28c..e02a28266 100644
--- a/src/get_ld.c
+++ b/src/get_ld.c
@@ -46,7 +46,12 @@ mpfr_get_ld (mpfr_srcptr x, mpfr_rnd_t rnd_mode)
          all subsequent operations are exact (this avoids double rounding
          problems) */
       mpfr_init2 (y, MPFR_LDBL_MANT_DIG);
-      mpfr_init2 (z, IEEE_DBL_MANT_DIG);
+      mpfr_init2 (z, MPFR_LDBL_MANT_DIG);
+      /* Note about the precision of z: even though IEEE_DBL_MANT_DIG is
+         sufficient, z has been set to the same precision as y so that
+         the mpfr_sub below calls mpfr_sub1sp, which is faster than the
+         generic subtraction, even in this particular case (from tests
+         done by Patrick Pelissier on a 64-bit Core2 Duo against r7285). */
 
       mpfr_set (y, x, rnd_mode);
       sh = MPFR_GET_EXP (y);