[libc-commits] [libc] [libc][math] Improve hypotf performance. (PR #186627)

[via libc-commits]

================
@@ -25,79 +25,96 @@ namespace math {
 LIBC_INLINE float hypotf(float x, float y) {
   using DoubleBits = fputil::FPBits<double>;
   using FPBits = fputil::FPBits<float>;
+  using fputil::DoubleDouble;
 
-  FPBits x_abs = FPBits(x).abs();
-  FPBits y_abs = FPBits(y).abs();
+  uint32_t x_a = FPBits(x).uintval() & 0x7fff'ffff;
+  uint32_t y_a = FPBits(y).uintval() & 0x7fff'ffff;
 
-  bool x_abs_larger = x_abs.uintval() >= y_abs.uintval();
-
-  FPBits a_bits = x_abs_larger ? x_abs : y_abs;
-  FPBits b_bits = x_abs_larger ? y_abs : x_abs;
-
-  uint32_t a_u = a_bits.uintval();
-  uint32_t b_u = b_bits.uintval();
-
-  // Note: replacing `a_u >= FPBits::EXP_MASK` with `a_bits.is_inf_or_nan()`
+  // Note: replacing `x_a >= FPBits::EXP_MASK` with `x_bits.is_inf_or_nan()`
   // generates extra exponent bit masking instructions on x86-64.
-  if (LIBC_UNLIKELY(a_u >= FPBits::EXP_MASK)) {
+  if (LIBC_UNLIKELY(x_a >= FPBits::EXP_MASK || y_a >= FPBits::EXP_MASK)) {
     // x or y is inf or nan
-    if (a_bits.is_signaling_nan() || b_bits.is_signaling_nan()) {
+    FPBits x_bits(x);
+    FPBits y_bits(y);
+    if (x_bits.is_signaling_nan() || y_bits.is_signaling_nan()) {
       fputil::raise_except_if_required(FE_INVALID);
       return FPBits::quiet_nan().get_val();
     }
-    if (a_bits.is_inf() || b_bits.is_inf())
+    if (x_bits.is_inf() || y_bits.is_inf())
       return FPBits::inf().get_val();
-    return a_bits.get_val();
+    return x + y;
   }
 
-  if (LIBC_UNLIKELY(a_u - b_u >=
----------------
lntue wrote:

The throughput and latency of the main path is low enough.  Simply trading off between the performance of when a & b are close enough vs when b is negligible (which should be rarer).

https://github.com/llvm/llvm-project/pull/186627