[libc++] Add hack to allow ubsan to work w/o compiler-rt (__muloti4 is undefined)

[EricWF]

Summary

Using int128_t with UBSAN causes link errors unless compiler-rt is providing the runtime library.
Specifically ubsan generates calls to __muloti4 but libgcc doesn't provide a definition.

In order to avoid this, and allow users to continue using sanitized versions of libc++, this patch introduces a hack.
It adds a cribbed version of the compiler-rt builtin to the libc++ filesystem sources.

I don't think this approach will work in the long run, but it seems OK for now.

Also see:

https://bugs.llvm.org/show_bug.cgi?id=30643
https://bugs.llvm.org/show_bug.cgi?id=16404

[echristo]

Bit of a hack, but I'm ok with it.

src/experimental/filesystem/int128_builtins.cpp

+/*===-- int128_builtins.cpp - Implement __muloti4 --------------------------===
+ *
+ *                     The LLVM Compiler Infrastructure
+ *
+ * This file is dual licensed under the MIT and the University of Illinois Open
+ * Source Licenses. See LICENSE.TXT for details.
+ *
+ * ===----------------------------------------------------------------------===
+ *
+ * This file implements __muloti4, and is stolen from the compiler_rt library.
+ *
+ * FIXME: we steal and re-compile it into filesystem, which uses __int128_t,
+ * and requires this builtin when sanitized. See llvm.org/PR30643
+ *
+ * ===----------------------------------------------------------------------===
+ */
+#include "__config"
+#include "climits"
+
+#ifndef _LIBCPP_HAS_NO_INT128
+
+__attribute__((no_sanitize("undefined"))) extern "C" __int128_t
+__muloti4(__int128_t a, __int128_t b, int* overflow) {
+  const int N = (int)(sizeof(__int128_t) * CHAR_BIT);
+  const __int128_t MIN = (__int128_t)1 << (N - 1);
+  const __int128_t MAX = ~MIN;
+  *overflow = 0;
+  __int128_t result = a * b;
+  if (a == MIN) {
+    if (b != 0 && b != 1)
+      *overflow = 1;
+    return result;
+  }
+  if (b == MIN) {
+    if (a != 0 && a != 1)
+      *overflow = 1;
+    return result;
+  }
+  __int128_t sa = a >> (N - 1);
+  __int128_t abs_a = (a ^ sa) - sa;
+  __int128_t sb = b >> (N - 1);
+  __int128_t abs_b = (b ^ sb) - sb;
+  if (abs_a < 2 || abs_b < 2)
+    return result;
+  if (sa == sb) {
+    if (abs_a > MAX / abs_b)
+      *overflow = 1;
+  } else {
+    if (abs_a > MIN / -abs_b)
+      *overflow = 1;
+  }
+  return result;
+}
+
+#endif