[libc-commits] [libc] [libc][math] Fix `tanpif` and `tanpif16` error handling (PR #235296)

[via libc-commits]

================
@@ -53,13 +53,24 @@ LIBC_INLINE float16 tanpif16(float16 x) {
   if (LIBC_UNLIKELY(x_abs == 0U))
     return x;
 
+  // tanpi(x) underflows for 0 < |x| <= 0x1.45p-16.
+  if (LIBC_UNLIKELY(x_abs <= 0x0145U)) {
+    fputil::set_errno_if_required(ERANGE);
+    fputil::raise_except_if_required(FE_UNDERFLOW | FE_INEXACT);
+  }
+
   if (LIBC_UNLIKELY(fputil::is_integer(x))) {
     // Preserve the input sign for even integers and flip it for odd integers.
     Sign sign =
         fputil::is_odd_integer(x) ? xbits.sign().negate() : xbits.sign();
     return FPBits::zero(sign).get_val();
   }
 
+  // Odd multiples of 1/4 have exact results of +1 or -1.
+  float xf = x;
+  if (LIBC_UNLIKELY(fputil::is_odd_integer(4.0f * xf)))
+    return fputil::is_odd_integer(2.0f * xf - 0.5f) ? -1.0f16 : 1.0f16;
----------------
lntue wrote:

use `multiply_add`.

https://github.com/llvm/llvm-project/pull/235296