speed up Float16 conversions a bit

base/float.jl

@@ -144,17 +144,17 @@ function Float16(val::Float32)
         return reinterpret(Float16, t ⊻ ((f >> 0xd) % UInt16))
     end
     i = (f >> 23) & 0x1ff + 1
-    sh = shifttable[i]
+    @inbounds sh = shifttable[i]
     f &= 0x007fffff
-    h::UInt16 = basetable[i] + (f >> sh)
+    @inbounds h = (basetable[i] + (f >> sh)) % UInt16
     # round
     # NOTE: we maybe should ignore NaNs here, but the payload is
     # getting truncated anyway so "rounding" it might not matter
     nextbit = (f >> (sh-1)) & 1
     if nextbit != 0
         # Round halfway to even or check lower bits
         if h&1 == 1 || (f & ((1<<(sh-1))-1)) != 0
-            h += 1
+            h += UInt16(1)
         end
     end
     reinterpret(Float16, h)
@@ -179,7 +179,7 @@ function Float32(val::Float16)
                 bit = bit >> 1
             end
             sign = sign << 31
-            exp = (-14 - n_bit + 127) << 23
+            exp = ((-14 - n_bit + 127) << 23) % UInt32
             sig = ((sig & (~bit)) << n_bit) << (23 - 10)
             ret = sign | exp | sig
         end
@@ -195,7 +195,7 @@ function Float32(val::Float16)
         end
     else
         sign = sign << 31
-        exp  = (exp - 15 + 127) << 23
+        exp  = ((exp - 15 + 127) << 23) % UInt32
         sig  = sig << (23 - 10)
         ret = sign | exp | sig
     end