LLVM cannot constant fold memchr with constant size

[philnik777]

void* test1(char* first, char val) {
  return __builtin_memchr(first, val, 2);
}

always generates a call to memchr, even though it could just check the two chars. I can't really see a case where calling memchr is faster than just checking the two chars. For larger N it would also be interesting to use vector comparisons for this instead of calling a function.

[nikic]

This needs an expansion to branchy code, so it could go into AggressiveInstCombine.