[Builtins] Implement half-precision conversions.

Mostly uninteresting, except:
- in __extendXfYf2, when checking if the number is normal, the old
  code relied on the unsignedness of src_rep_t, which is a problem
  when sizeof(src_rep_t) < sizeof(int): the result gets promoted to
  int, the signedness of which breaks the comparison.
  I added an explicit cast; it shouldn't affect other types.
- we can't pass __fp16, so src_t and src_rep_t are the same.
- the gnu_*_ieee symbols are simply duplicated definitions, as aliases
  are problematic on mach-o (where only weak aliases are supported;
  that's not what we want).

Differential Revision: http://reviews.llvm.org/D9693

llvm-svn: 237161
diff --git a/compiler-rt/lib/builtins/fp_extend_impl.inc b/compiler-rt/lib/builtins/fp_extend_impl.inc
index f6953ff..edcfa8d 100644
--- a/compiler-rt/lib/builtins/fp_extend_impl.inc
+++ b/compiler-rt/lib/builtins/fp_extend_impl.inc
@@ -66,7 +66,9 @@
     const src_rep_t sign = aRep & srcSignMask;
     dst_rep_t absResult;
 
-    if (aAbs - srcMinNormal < srcInfinity - srcMinNormal) {
+    // If sizeof(src_rep_t) < sizeof(int), the subtraction result is promoted
+    // to (signed) int.  To avoid that, explicitly cast to src_rep_t.
+    if ((src_rep_t)(aAbs - srcMinNormal) < srcInfinity - srcMinNormal) {
         // a is a normal number.
         // Extend to the destination type by shifting the significand and
         // exponent into the proper position and rebiasing the exponent.