[libc-commits] [libc] [libc][stdfix] Implement `countlsfx` functions (PR #114318)

[via libc-commits]

================
@@ -142,6 +160,25 @@ template <typename T> LIBC_INLINE constexpr T abs(T x) {
   }
 }
 
+template <typename T>
+LIBC_INLINE constexpr cpp::enable_if_t<cpp::is_fixed_point_v<T>, int>
+countls(T x) {
+  using FXRep = FXRep<T>;
+  using BitType = typename FXRep::StorageType;
+  constexpr int CONTAIN_LEN = cpp::numeric_limits<BitType>::digits;
+  constexpr int PADDING_LEN = CONTAIN_LEN - FXRep::VALUE_LEN;
+
+  if constexpr (FXRep::SIGN_LEN != 0) {
+    if (x < 0) {
+      x = -(x + FXRep::EPS());
+    }
+  }
+
+  FXBits<T> fxbits(x);
+  BitType value_bits = fxbits.get_value_bits();
----------------
duncpro wrote:

So rename `get_value_bits` to `get_bits` and
inside `get_bits` we `& TOTAL_MASK` instead of `& VALUE_MASK`.

The idea is...`TOTAL_MASK` is actually `StorageType::MAX` in most cases, and so clang can easily see its useless and optimize away.

Whereas it can't as easily get optimized away if we use `VALUE_MASK` because we discard the sign bit.

https://github.com/llvm/llvm-project/pull/114318