Workaround for dtoa.cbb-10.3-vicentiu-dtoa

There is no obvious reason why multiplying by 1.0 helps solve a
precision problem on gcc-7 or clang-5, but it does work. Casting to
ulong however makes the problem come up again. Windows warns about loss
of precision, so to quiet warnings until an appropriate fix is
implemented, ifdef 2 versions of the code.

1 files changed, 11 insertions, 1 deletions

diff --git a/strings/dtoa.c b/strings/dtoa.c
index 7b560227026..f27a91c3920 100644
--- a/strings/dtoa.c
+++ b/strings/dtoa.c
@@ -1290,7 +1290,17 @@ static double ratio(Bigint *a, Bigint *b)
   dval(&db)= b2d(b, &kb);
   k= ka - kb + 32*(a->wds - b->wds);
   if (k > 0)
-    word0(&da)+= (ULong)(k*Exp_msk1 * 1.0);
+#ifdef _WIN32
+    word0(&da)+= k*Exp_msk1;
+#else
+  /* 
+     TODO(cvicentiu)
+     This is a temporary fix for a possible clang-5 / gcc-7 bug. This
+     should not be required but it makes double results in the end
+     be predictable.
+   */
+    word0(&da)+= k*Exp_msk1 * 1.0;
+#endif
   else
   {
     k= -k;