[libc-commits] [libc] [libc][math][c23] Add f16sqrt{, l, f128} C23 math functions (PR #96642)

[via libc-commits]

================
@@ -146,91 +144,19 @@ sqrt(InType x) {
       }
 
       // We compute one more iteration in order to round correctly.
-      bool lsb = (y & (InStorageType(1) << EXTRA_FRACTION_LEN)) !=
-                 0;    // Least significant bit
-      bool rb = false; // Round bit
       r <<= 2;
-      InStorageType tmp = (y << 2) + 1;
+      y <<= 2;
+      InStorageType tmp = y + 1;
       if (r >= tmp) {
         r -= tmp;
-        rb = true;
-      }
-
-      bool sticky = false;
-
-      if constexpr (EXTRA_FRACTION_LEN > 0) {
-        sticky = rb || (y & EXTRA_FRACTION_MASK) != 0;
-        rb = (y & (InStorageType(1) << (EXTRA_FRACTION_LEN - 1))) != 0;
-      }
-
-      // Remove hidden bit and append the exponent field.
-      x_exp = ((x_exp >> 1) + OutFPBits::EXP_BIAS);
-
-      OutStorageType y_out = static_cast<OutStorageType>(
-          ((y - ONE) >> EXTRA_FRACTION_LEN) |
-          (static_cast<OutStorageType>(x_exp) << OutFPBits::FRACTION_LEN));
-
-      if constexpr (EXTRA_FRACTION_LEN > 0) {
-        if (x_exp >= OutFPBits::MAX_BIASED_EXPONENT) {
-          switch (quick_get_round()) {
-          case FE_TONEAREST:
-          case FE_UPWARD:
-            return OutFPBits::inf().get_val();
-          default:
-            return OutFPBits::max_normal().get_val();
-          }
-        }
-
-        if (x_exp <
-            -OutFPBits::EXP_BIAS - OutFPBits::SIG_LEN + EXTRA_FRACTION_LEN) {
-          switch (quick_get_round()) {
-          case FE_UPWARD:
-            return OutFPBits::min_subnormal().get_val();
-          default:
-            return OutType(0.0);
-          }
-        }
-
-        if (x_exp <= 0) {
-          int underflow_extra_fraction_len = EXTRA_FRACTION_LEN - x_exp + 1;
-          InStorageType underflow_extra_fraction_mask =
-              (InStorageType(1) << underflow_extra_fraction_len) - 1;
-
-          rb = (y & (InStorageType(1) << (underflow_extra_fraction_len - 1))) !=
-               0;
-          OutStorageType subnormal_mant =
-              static_cast<OutStorageType>(y >> underflow_extra_fraction_len);
-          lsb = (subnormal_mant & 1) != 0;
-          sticky = sticky || (y & underflow_extra_fraction_mask) != 0;
-
-          switch (quick_get_round()) {
-          case FE_TONEAREST:
-            if (rb && (lsb || sticky))
-              ++subnormal_mant;
-            break;
-          case FE_UPWARD:
-            if (rb || sticky)
-              ++subnormal_mant;
-            break;
-          }
-
-          return cpp::bit_cast<OutType>(subnormal_mant);
-        }
-      }
-
-      switch (quick_get_round()) {
-      case FE_TONEAREST:
-        // Round to nearest, ties to even
-        if (rb && (lsb || (r != 0)))
-          ++y_out;
-        break;
-      case FE_UPWARD:
-        if (rb || (r != 0) || sticky)
-          ++y_out;
-        break;
+        // Rounding bit.
+        y |= 1 << 1;
----------------
lntue wrote:

Just put `y |= 2` is fine.
Also, can you add a TODO on the main for loop to reduce it to `OutFPBits::FRACTION_LEN + 2` or 3?

https://github.com/llvm/llvm-project/pull/96642