[libc-commits] [libc] [libc][math][c23] Improve rsqrtf16() function (PR #160639)

[via libc-commits]

================
@@ -76,6 +76,76 @@ LIBC_INLINE static constexpr float16 rsqrtf16(float16 x) {
   }
 
   return fputil::cast<float16>(result);
+
+#else
+  float xf = fputil::cast<float>(x);
+
+  int exponent = 0;
+  float mantissa = fputil::frexp(xf, exponent);
----------------
lntue wrote:

Use `FPBits` and direct bit manipulations instead of `fputil::frexp` and `fputil::ldexp`, because those 2 functions do a lot more work (w.r.t floating-point exceptions and errnos) than just extract and normalize parts of the floating-point.

https://github.com/llvm/llvm-project/pull/160639