add regression test for #285 coshl tiny inputs

Covers ld128 coshl(2^-k) for k in [57,128]; the pre-fix code returned
1+tiny instead of exactly 1.0 for k>71. Skipped where long double is
not 128-bit (LDBL_MANT_DIG < 100).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: ViralBShah

test/regression/test-285.c

@@ -0,0 +1,52 @@
+/*
+ * Regression test for issue #285: ld128 coshl returned 1+x instead of 1 for
+ * tiny |x|.
+ * https://github.com/JuliaMath/openlibm/issues/285
+ *
+ * The early-exit branch in ld128/e_coshl.c returned `1 + expm1l(x)` (~= 1+x)
+ * for tiny representable x instead of exactly 1.0. cosh(x) = 1 + x^2/2 + O(x^4);
+ * for these inputs the x^2/2 term is below half an ulp at 1.0 (2^-113 in
+ * 113-bit precision), so the correctly rounded result is exactly 1.0.
+ *
+ * This only manifests where long double is 128-bit (LDBL_MANT_DIG == 113),
+ * e.g. aarch64/riscv64 Linux. Elsewhere (80-bit x87, 64-bit double) the test
+ * is skipped.
+ */
+#include <float.h>
+
+#include "regress-util.h"
+
+/*
+ * The fix is in ld128/e_coshl.c, which openlibm compiles only on (linux)
+ * aarch64.  Restrict the test to that target at compile time so that on every
+ * other arch it becomes a bare skip with no reference to coshl (which may be
+ * absent from the library and would otherwise break the link).
+ */
+#if defined(__aarch64__) && LDBL_MANT_DIG == 113
+
+#include <math.h>
+
+int
+main(void)
+{
+	int k;
+
+	/*
+	 * For k in [57,128], coshl(2^-k) rounds to exactly 1.0L in 113-bit
+	 * precision. The pre-fix code returned 1+tiny for k > 71.
+	 */
+	for (k = 57; k <= 128; k++)
+		CHECK(coshl(ldexpl(1.0L, -k)) == 1.0L);
+
+	return 0;
+}
+
+#else
+
+int
+main(void)
+{
+	return REGRESS_SKIP;
+}
+
+#endif