[compiler-rt] Add extendhfxf2 into compiler rt (PR #113897)

[Alexander Richardson via llvm-commits]

================
@@ -0,0 +1,73 @@
+// RUN: %clang_builtins %s %librt -o %t && %run %t
+// REQUIRES: librt_has_extendhfxf2
+
+#include <limits.h>
+#include <math.h> // for isnan, isinf
+#include <stdio.h>
+
+#include "int_lib.h"
+
+#if HAS_80_BIT_LONG_DOUBLE && defined(COMPILER_RT_HAS_FLOAT16)
+
+long double __extendhfxf2(_Float16 f);
+
+int test_extendhfxf2(_Float16 a, long double expected) {
+  long double x = __extendhfxf2(a);
+  __uint16_t *b = (void *)&a;
+  int ret = !((isnan(x) && isnan(expected)) || x == expected);
+  if (ret) {
+    printf("error in test__extendhfxf2(%#.4x) = %.20Lf, "
+           "expected %.20Lf\n",
+           *b, x, expected);
+  }
+  return ret;
+}
+
+char assumption_1[sizeof(_Float16) * CHAR_BIT == 16] = {0};
+
+int main() {
+  // Small positive value
+  if (test_extendhfxf2(0.09997558593750000000f, 0.09997558593750000000L))
+    return 1;
+
+  // Small negative value
+  if (test_extendhfxf2(-0.09997558593750000000f, -0.09997558593750000000L))
+    return 1;
+
+  // Zero
+  if (test_extendhfxf2(0.0f, 0.0L))
+    return 1;
+
+  // Smallest positive non-zero value
+  if (test_extendhfxf2(0x1p-16f, 0x1p-16L))
+    return 1;
+
+  // Smallest negative non-zero value
+  if (test_extendhfxf2(-0x1p-16f, -0x1p-16L))
+    return 1;
+
+  // Positive infinity
+  if (test_extendhfxf2(__builtin_huge_valf16(), __builtin_huge_valf64x()))
+    return 1;
+
+  // Negative infinity
+  if (test_extendhfxf2(-__builtin_huge_valf16(),
+                       (long double)-__builtin_huge_valf64x()))
+    return 1;
+
+  // NaN
+  if (test_extendhfxf2(__builtin_nanf16(""),
+                       (long double)__builtin_nanf64x("")))
+    return 1;
----------------
arichardson wrote:

The ones that support float16 should provide this function.

https://github.com/llvm/llvm-project/pull/113897